Tennessee Eastman Benchmark Process Research Articles

Fault Detection Analysis Before and After Dynamic Model Reforming on the Benchmark Tennessee Eastman Process

To make the simulation object closer to the actual chemical process, the author improved the model of Tennessee Eastman (TE) process in their previous work. In this paper, based on above improved model, to further analyse its effect on the fault detection performance, a research on fault data of the improved model and original model is done, based on principle component analysis (PCA), the detection rate of Hotelling's T2 statistic, Q statistic and support vector machine (SVM) integrated particle swarm optimization (PSO) approach are used to reflect their detection performance on the two models. From the detection rates, the detection performance gets worse when detecting the fault of the improved model. The analysis indicates that when the detection methods are used to detect the faults in actual chemical process, the detection performance will be influenced and may not be as effective as described in literature.

Generalized grouped contributions for hierarchical fault diagnosis with group Lasso

In process industries, it is necessary to conduct fault diagnosis after abnormality is found, with the aim to identify root cause variables and further provide instructive information for maintenance. Contribution plots along with multivariate statistical process monitoring are standard tools towards this goal, which, however, suffer from the smearing effect and high diagnostic complexity on large-scale processes. In fact, process variables tend to be naturally grouped, and in this work, a novel fault identification strategy based on group Lasso penalty along with a hierarchical fault diagnosis scheme is proposed by leveraging group information among variables. By introducing the group Lasso as a regularization approach, groups of irrelevant variables tend to yield exactly zero contributions collectively, which help find the exact root cause, alleviate the smearing effect, and furnish clear diagnostic information for process practitioners. For online computational convenience, an efficient numerical solution strategy is also presented. Besides, it turns out that the proposed approach also applies to dynamic monitoring models with lagged measurements augmented, thereby enjoying widespread generality. Its effectiveness is evaluated on both the Tennessee Eastman benchmark process and a pilot-scale experiment apparatus.

Distributed Dynamic Process Monitoring Based on Minimal Redundancy Maximal Relevance Variable Selection and Bayesian Inference

The conventional distributed modeling strategy generally includes all the process variables in large-scale process monitoring, thus submerging the local fault information. Meanwhile, fault diagnosis issues in the aforementioned process are also worth studying. To make up the deficiencies of the general distributed method, this brief presents an advantageously distributed fault monitoring and diagnosis scheme for large-scale dynamic processes. First, the optimal variable subblock is established by the minimal redundancy maximal relevance (mRMR) algorithm for each measured variable individually. Second, the corresponding monitoring model is set up in every subblock, respectively. Finally, all the distributed multiblocks’ results are combined to form an integrated decision through the Bayesian inference. The proposed scheme considers not only the interpretation of the correlation of variables but also the redundancy between them in the block division, which better characterizes the dynamic relationships among variables, thereby facilitating the monitoring performance significantly. After that, a new contribution plot method based on subblock analysis is applied for fault isolation. Both a real-world industrial process and the Tennessee Eastman benchmark process validate that the novel monitoring strategy is more efficient than the existing state-of-the-art approaches.

Multirate Dynamic Process Monitoring Based on Multirate Linear Gaussian State-Space Model

Multivariate statistical process monitoring (MSPM) has been widely used in modern industries and most of traditional MSPM methods are developed using uniformly sampled measurements. However, process variables are often sampled with different rates in practical industries. On the other hand, most of the industries are dynamic processes in which the measurements are highly autocorrelated. Thus, it is difficult to build a dynamic process model with incomplete data sets in multirate processes. In this paper, a multirate linear Gaussian state-space model is exploited to deal with the above issues. Both the offline model training and online process monitoring schemes are developed in the present of incomplete multirate process data sets. The proposed method is validated through a numerical example and the Tennessee Eastman benchmark process. Note to Practitioners —Process monitoring is essential to industrial process maintenance and safety. Typical monitoring methods focus on uniformly sampled variables which are rarely satisfied in practical industries. On the other hand, most industries are dynamic processes which bring more difficulties to process monitoring with multirate variables. This paper proposed a new scheme for multirate dynamic process monitoring based on the multirate linear Gaussian state-space model. To fully implement this approach: 1) the multirate process data need to be well divided into multiblocks. Each of the block should contain at least one variable; 2) the amount of latent variable’s dimension needs to be properly determined; and 3) the sampling rates of the process cannot be extremely distinguished otherwise will lead to a poor model performance. Two cases studied on both numerical example and Tennessee Eastman benchmark are given to validate the efficiency of the proposed algorithm on multirate dynamic process monitoring.

Mutual information-based sparse multiblock dissimilarity method for incipient fault detection and diagnosis in plant-wide process

Multiblock methods have attracted much attention in monitoring of plant-wide processes. However, most recent works only provide rough division results and do not take the data distribution changes among the process into consideration. To solve these defects, a new multiblock monitoring scheme that integrates a mutual information (MI)-based sparsification method, dissimilarity analysis (DISSIM) method and support vector data description (SVDD) is proposed in this paper. This method makes use of the complex relations between variables and the connections between sub-blocks in process decomposition to produce easy-to-interpret sub-blocks related to the process mechanism. Then DISSIM method is applied in each sub-block for distribution monitoring. The multiblock DISSIM strategy can deal with the local behaviors and distribution changes, improving its sensitivity to incipient changes in plant-wide process. Finally, results in all blocks are combined with SVDD to provide a final decision, and a diagnosis method is proposed for fault diagnosis. Case studies upon a numerical case and Tennessee Eastman (TE) benchmark process demonstrate the effectiveness of our method.

Fault Classification in High-Dimensional Complex Processes Using Semi-Supervised Deep Convolutional Generative Models

In complex industrial processes, process fault detection and classification constitute an important task for reducing production costs and improving product quality. Most existing methods for fault classification assume that sufficient labeled data are available for training. However, label acquisition is costly and laborious in practice, whereas abundant unlabeled data are often available. To make effective use of a large amount of unlabeled data for fault classification, we propose in this article a new approach using semi-supervised deep generative models, allowing the complex relationship between high-dimensional process data and the process status to be modeled. In particular, to consider the temporal correlation and intervariable correlation in multivariate time series process data collected from multiple sensors, we propose two semi-supervised deep generative models incorporating convolutional neural networks. The proposed models are assessed on data from the Tennessee Eastman benchmark process. The results demonstrate the superior performances of the proposed models compared with competing methods.

A single Bayesian network classifier for monitoring with unknown classes

In this paper, the Conditional Gaussian Networks (CGNs), a form of Bayesian Networks (BN), are used as a statistical process monitoring approach to detect and diagnose faults. The proposed approach improves the structure of Bayesian networks and generalizes a few results regarding statistical tests and the use of an exclusion criterion. The proposed framework is evaluated using data from the benchmark Tennessee Eastman Process (TEP) with various scenarios.

Manifold-Preserving Sparse Graph-Based Ensemble FDA for Industrial Label-Noise Fault Classification

For the fault classification of chemical industries, the typical Fisher discriminant analysis (FDA) model requires that all the training samples should be correctly labeled. Actually, training samples tend to be polluted by mislabeled samples for some unavoidable reasons, and thus the performance of the FDA may be severely affected. However, in the chemical process monitoring community, there is little attention to fault classification with label-noise in training samples. To settle this open issue, a manifold-preserving sparse graph (MPSG)-based ensemble FDA model is initially proposed in this paper. First, under the condition of maintaining the underlying manifold structure of training samples, an MPSG is utilized to filter some mislabeled samples. Second, to improve model robustness to the left mislabeled samples, Bagging based on FDA is used to construct several subclassifiers, which are combined to form a robust classifier. Experiments on the benchmark Tennessee Eastman process and a real industrial air separation unit demonstrate the effectiveness and the superiority of the proposed model.

K-means Bayes algorithm for imbalanced fault classification and big data application

Fault classification is an important part of process monitoring in industrial processes. Most conventional fault classification methods are under the assumption that the amount of data in different classes are similar. However, in practice, most of the data collected from industrial process are normal data (majority) and only a few of them are fault data (minority). In other words, fault classification can be seen as an imbalanced data classification problem, which has not been considered in this area to date. In this paper, a K-means Bayes algorithm is proposed to deal with the imbalanced fault classification problem. After that, a MapReduce approach is further introduced to implement the method for fault classification in the big data case. Effectiveness of the proposed method is verified through experiments based on Tennessee Eastman (TE) benchmark process.

Decentralized monitoring for large‐scale process using copula‐correlation analysis and Bayesian inference–based multiblock principal component analysis

AbstractDue to the massive monitored samples and their features such as strong coupling and time delay in large‐scale industrial process, the decentralized monitoring methods, which divide the variables into several blocks and perform local monitoring in each subblock, have been brought up to handle the complex characteristics. However, most of the existing decentralized methods utilize the correlation degree among variables for block division, while the correlation patterns among variables are ignored. The missing of correlation pattern will influence the monitoring performance. To address this problem, the copula‐correlation analysis, which considers both the correlation degree and the correlation pattern, is adopted for block division; then the principal component analysis (PCA)–based method is used for subblock monitoring, and the Bayesian inference is introduced to achieve decision fusion of fault detection. The superiority and effectiveness of the proposed method are illustrated through comparison studies on a numerical example and the Tennessee Eastman (TE) benchmark process.

Fault Diagnosis of Complex Processes Using Sparse Kernel Local Fisher Discriminant Analysis.

As an outstanding discriminant analysis technique, Fisher discriminant analysis (FDA) gained extensive attention in supervised dimensionality reduction and fault diagnosis fields. However, it typically ignores the multimodality within the measured data, which may cause infeasibility in practice. In addition, it generally incorporates all process variables without emphasizing the key faulty ones when modeling the complex process, thus leading to degraded fault classification capability and poor model interpretability. To ease the above two drawbacks of conventional FDA, this brief presents an advantageously sparse local FDA (SLFDA) model, it first preserves the within-class multimodality by introducing local weighting factors into scatter matrix. Then, the responsible faulty variables are identified automatically through the elastic net algorithm, and the current optimization problem is subsequently settled through the feasible gradient direction method. Since then, the local data structure characteristics are exploited from both the sample dimension and variable dimension so that the fault diagnosis performance and model interpretability are significantly enhanced. In addition, we naturally extend SLFDA model to nonlinear variant (i.e., sparse kernel local FDA) by the kernel trick, which is substantially more resistant to strong nonlinearity. The simulation studies on Tennessee Eastman (TE) benchmark process and real-world diesel engine working process both validate that the novel diagnosis strategy is more accurate and reliable than the existing state-of-the-art methods.

Generating optimal overlapping subsystems for distributed statistical fault detection subject to constraints

Fast and accurate detection of faults is essential for the safe and cost-effective operation of chemical plants. Large scale process systems can potentially be affected by a wide variety of faults. An effective approach for detecting faults in such systems is to first divide or decompose the system's sensors into subsystems and to then implement a statistical monitoring method in each of the subsystems. The detection performance of such distributed multivariate statistical process monitoring methods depends strongly on the decomposition of the system's sensors into subsystems. In our previous work, we developed a simulation optimization method, called Performance Driven Agglomerative Clustering, which uses a greedy search strategy, based on clustering algorithms from graph theory, to find a decomposition of the system's sensors into subsystems for which the detection performance of a distributed monitoring method is near optimal. It is possible that the detection performance can be further improved by allowing a sensor to be part of multiple subsystems in the decomposition. User defined requirements may also place constraints on the decomposition of the sensors into subsystems. In this work, we propose the Extended Performance Driven Agglomerative Clustering method which allows to incorporate constraints and sensors to be allocated to multiple subsystems. To demonstrate its effectiveness, the proposed method is applied to the benchmark Tennessee Eastman Process.

Global-and-local-structure-based neural network for fault detection

A novel statistical fault detection method, called the global-and-local-structure-based neural network (GLSNN), is proposed for fault detection. GLSNN is a nonlinear data-driven process monitoring technique through preserving both global and local structures of normal process data. GLSNN is characterized by adaptively training a neural network which takes both the global variance information and the local geometrical structure into consideration. GLSNN is designed to extract the meaningful low-dimensional features from original high-dimensional process data. After nonlinear feature extraction, Hotelling T2 statistic and the squared prediction error (SPE) statistic are adopted for online fault detection. The merits of the proposed GLSNN method are demonstrated by both theoretical analysis and case studies on the Tennessee Eastman (TE) benchmark process. Extensive experimental results show the superiority of GLSNN in terms of missed detection rate (MDR) and false alarm rate (FAR). The source code of GLSNN can be found in https://github.com/htzhaoecust/glsnn.

Change point and fault detection using Kantorovich Distance

The automation of real-time process monitoring in the industry is an ongoing challenge. Chief among the objectives of monitoring are change point detection and the detection of faults in process variables and sensor measurements. In this paper, we propose a novel algorithm for change point and fault detection using Kantorovich Distance (KD), a metric induced from optimal mass transport theory. To evaluate the performance of the proposed method, we first evaluate the change point detection capability of the KD metric for data sampled from various probability distributions. Next, the fault detection performance of the KD metric is evaluated for three cases of faults – sustained bias, drift, and multiple intermittent biases – and contrasted against that of the traditional PCA-based metrics, Q and T2 statistics. The algorithm is tested on several case studies including a synthetic data, a simulated continuous stirred tank heater system and the benchmark Tennessee Eastman process. The results obtained showcase the superiority of the proposed algorithm over the conventional scheme.

Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning

This study presents an ensemble monitoring strategy based on slow feature analysis (SFA) model of multi-subspace partitioning for dynamic large-scale process. SFA can effectively extract the various dynamics of process data, where the relationship between process data and slow features (SFs) can be revealed by transformation matrix. The similar projecting directions represent similar importance of variables, and corresponding latent variables (LVs) will show similar monitoring behavior. Several LV subspaces are obtained by dividing the transformation vectors with higher similarity into the same sub-block automatically based on the defined process variable related index and hierarchical clustering, which can avoid the problems of information loss and the selection of SFs. Then, the S2 statistics constructed in each subspaces are integrated by support vector data description to show an intuitive detection results. Experiments on Tennessee Eastman benchmark process and wastewater treatment process have validated the proposed strategy’s effectiveness and excellence.

Data-driven digital twin technology for optimized control in process systems

Due to the installation of various apparatus in process industries, both factors of complex structures and severe operating conditions could result in higher accident frequencies and maintenance challenges. Given the importance of security in process systems, this paper presents a data-driven digital twin system for automatic process applications by integrating virtual modeling, process monitoring, diagnosis, and optimized control into a cooperative architecture. For unknown model parameters, the adaptive system identification is proposed to model closed-loop virtual systems and residual signals with fault-free case data. Performance indices are improved to make the design of robust monitoring and diagnosis system to identify the apparatus status. Soft-sensor, parameterization control, and model-matching reconfiguration are ameliorated and incorporated into the optimized control configuration to guarantee stable and safe control performance under apparatus faults. The effectiveness and performance of the proposed digital twin system are evaluated by using different simulations on the Tennessee Eastman benchmark process in the presence of realistic fault scenarios.

A Distributed Canonical Correlation Analysis-Based Fault Detection Method for Plant-Wide Process Monitoring

In this paper, a new data-driven fault detection method based on distributed canonical correlation analysis (D-CCA) is proposed to address the plant-wide process monitoring problem. This paper focuses on the distributed plant-wide processes. The core of the proposed method is to reduce uncertainties using correlation information from the neighboring nodes. Furthermore, the cost of the data transmission between network nodes is also reduced by the D-CCA algorithm. When the proposed method and the existing methods are compared using the Tennessee Eastman benchmark process, the false alarm rate, fault detection rate, and the detection delay are comparable. This suggests that the proposed method is feasible.

SVM-tree and SVM-forest algorithms for imbalanced fault classification in industrial processes

Fault classification plays a central role in process monitoring and fault diagnosis in complex industrial processes. Plenty of fault classification methods have been proposed under the assumption that the sizes of different fault classes are similar. However, in practical industrial processes, it is a common case that large amounts of normal data (majority) and only few fault data (minority) are collected. In other words, most existing fault classification problems were carried out under the imbalanced data scenario, which will lead to a restricted performance of traditional classification algorithms. In this paper, a K-means based SVM-tree algorithm is proposed to deal with the nonlinear multiple-classification problem under the situation of imbalance data. Meanwhile, a SVM-forest scheme is further developed for sensitive data selection and performance enhancement when the imbalance degree is larger among different classes. Effectiveness of the proposed method is verified through the Tennessee Eastman (TE) benchmark process.

An Intelligent Fault Diagnosis Method Using GRU Neural Network towards Sequential Data in Dynamic Processes

Intelligent fault diagnosis is a promising tool to deal with industrial big data due to its ability in rapidly and efficiently processing collected signals and providing accurate diagnosis results. In traditional static intelligent diagnosis methods, however, the correlation between sequential data is neglected, and the features of raw data cannot be effectively extracted. Therefore, this paper proposes a three-stage fault diagnosis method based on a gate recurrent unit (GRU) network. The raw data is divided into several sequence units by first using a moving horizon as the input of GRU. In this way, we can intercept the sequence to get information as needed. Then, the GRU deep network is established through batch normalization (BN) algorithm to extract the dynamic feature from the sequence units effectively. Finally, the softmax regression is employed to classify faults based on dynamic features. Thus, the diagnosis result is obtained with a probabilistic explanation. Two chemical processes validate the proposed method: Tennessee Eastman (TE) benchmark process as well as para-xylene (PX) oxidation process. In the case of TE, the diagnosis results demonstrate the proposed method is superior to conventional methods. Furthermore, in the case of PX oxidation, the result shows that the proposed method also has an exceptional effect with a little historical data.

Mixtures of Probabilistic PCA With Common Structure Latent Bases for Process Monitoring

In this brief, we propose the mixtures of probabilistic principal component analyzers with latent bases having a common structure for modeling and monitoring multimodal processes. The proposed modeling framework attributes a joint distribution to each element of the latent bases across all the analyzers for bringing a consistent structure for the local models that correspond to various operating modes. Hierarchical prior distributions are attributed to regularize the parameters for obtaining sparse model structures. We employ the variational Bayesian expectation–maximization algorithm to train the model from the observed data. Faults are detected online if nonconformity of a data point to the developed model is identified. Furthermore, we identify faulty latent variables, and the process variables, which are significantly contributing to the faulty latent variables, are isolated by exploiting the unique structure of the model. We illustrate our proposed approach based on the simulations conducted on the Tennessee Eastman benchmark process.