Dynamic process fault monitoring based on neural network and PCA
Introduction
Over the past 20 years, the chemical industry has made a concerted effort to streamline operations. Their goal was simply to produce products as many as possible. Nowadays, as the market is highly competitive worldwide, production efficiency and product consistency become essential to success. Even though many chemical processes have been around for years and engineers have acquired lots of experience, many operational problems and inefficiencies still go undiagnosed for a prolonged period of time. Therefore, process monitoring and diagnosis are strongly required to produce the product and maintain the process equipment. For example, a heat exchanger that becomes fouled over a period of time may be unnoticed because it has no effect on the final product. Yet the incremental amount of the steam needs to be adjusted for fouling costs a significant amount of money. Process problems like this one should be monitored, detected and diagnosed. For most chemical processes, modern computers provide a system in which large amounts of data can be stored cheaply and efficiently. Currently, the process problems and inefficiencies are identified based on the historical data shown from the simple glorified strip charts or single variable statistics. Although the process data are accessible from any time period at the touch of a button, it remains difficult and time-consuming to fully understand how well the process is operated. It is difficult for everyone to find out the problem of examining time-sequenced data among all the variables because of the overwhelming amount of data, the existence of the multivariables and highly interacted nature of chemical processes. The knowledge a highly experienced person acquires about a process is seldom passed efficiently, if at all to his successor. This implies the strong need in developing help for operators who are confronted with hundreds of alarms coupled with conflicting indications. This is also particularly important for the automatic control systems since they are susceptible to faults that cause an unacceptable deterioration of the performance or even lead to dangerous situations.
Several techniques have been developed for monitoring and detection. These techniques can be broadly classified into three categories: model-based techniques, expert systems and pattern recognition [1], [2]. In the model-based approach, the actual behavior of the process to be supervised is compared with that of a nominal model driven by the same inputs. Faults can be detected or isolated by evaluating the difference between the estimated value of the model and the actual values of process variables. Some excellent survey research for overview of different aspects of this method is presented [3], [4], [5]. However, this method seems to be useful for limited applications [6] because the model-based approach is needed for governing equations that describe the process behavior as accurately as possible. The expert system, also called a knowledge-based system, is built upon some given facts and relations so as to make an induction for system behaviors. Examples of the expert system for fault detection can be found in Quantrille and Liu [7]. However, for many sophisticated chemical processes, if the fault related knowledge is not available or clear enough, it is very hard to develop an expert system. With the rapid progress on data process technology, pattern recognition has opened a new avenue in fault detection and diagnosis. Like the rule-based expert system, pattern recognition is based on the design of math-model free fault detection and diagnosis. From the concept of pattern mapping, the measurements and the identified fault model for each abnormal operation are needed in order to connect between patterns. Each pattern consists of measurements and corresponding fault models. The memories of the fault status are usually established via supervised or unsupervised training. When the patterns are established by a pattern mapping, or so-called retrieval process, any operating condition is assigned to a class or a label from a set of fault pattern into identifiable classes based on certain similar features [8]. Many well-known methods, such as artificial neural networks and fuzzy logic, belong to this category [9], [10], [11], [12], [13].
In recent years, chemometric techniques have been applied to monitoring and diagnosis in multivariable processes. Instead of using detailed mathematical models, they focus mostly on data-driven methods to extract the state of the system via applications of mathematical and statistical methods. The concept of monitoring and detection application is pretty close to that of the traditional statistical process control (SPC). The workhorse of SPC control charts, such as Shewhart chart, CUSUM (cumulative sum) and EWMA (exponentially weighted moving average), applies well to a monitoring process. CUSUM is a cumulative sum of the previous observations from the desired target. EWMA is actually just a smoothing algorithm, or a low-pass filter. The average value computed by EWMA can take noisy measurements and smooth them out (i.e. remove the highly fluctuating parts). The advantage of EWMA and CUSUM over Shewhart is particularly good for detecting small changes in mean. Shewart, however, gives a better performance in the case of large changes in the process mean. Details are readily obtained in many books on statistical quality control [14], [15]. EWMA and CUSUM charts for the multivariable problem are also developed [16], [17]. But those methods have limited use in a chemical production atmosphere since they do not comply with multivariable continuous processes with correlation among variables. As a result, it is very difficult to visualize the behavior without dimension reduction of the process variables. Several chemometric techniques, such as principal component analysis (PCA) and partial least squares, were developed and successfully applied to some industrial processes [18], [19], [20]. Unfortunately, these methods are only good for linear or closed-linear processes and they fail in nonlinear or dynamic processes.
For nonlinear systems, building the process model is extremely difficult in general. During the past few years, artificial neural networks (NN) were used to model nonlinear processes. Based on measurements of the process, a suitable NN can be trained to adapt the process behavior. Since NN requires little or no prior knowledge of systems, it provides an effective tool for dealing with nonlinearity because of its well-known approximation ability. This feature is particularly attracted to the fault diagnosis scheme due to the nonlinear nature of the problem. A general learning methodology for fault detection and diagnosis has been extensively studied, especially for steady-state systems [13] and for dynamic systems [12]. The former uses the network as a classifier of faults based on the process measurements and the latter as an alternative to the traditional model estimator. Some research has also applied NN to chemometric methods. Kramer (1992) proposed a nonlinear principal component analysis (NLPCA) based on the autoassociative neural network for uncovering linear and nonlinear relationship among the variables [21]. Dong and McAvoy (1996) developed a nonlinear principal component based on the principal curves and NN methods and applied it on batch processes [22]
The PCA based monitoring methods mentioned above are only developed for steady-state rather than dynamic relationships. That is, they implicitly assume that the measured variable at one time instant has not only serial independence within each variable series at past time instances but also statistical inter-independence between the different measured variable series at past time instances. Some researchers have combined two statistical process control methods to address this problem. For example, Wold [23] utilized EWMA on the score data from PCA, and Wachs and Lewin [24] constructed SSUM with PCA. Another way is to mimic the concept of the ARX time series model by forming the data matrix with the previous observations in each observation vector. This method that applies PCA to extracting the time-dependent relations in the measurements is referred to as Dynamic PCA (DPCA) [25]. It copes with linear system and it cannot be applied on the nonlinear chemical processes.
The purpose of this paper is to develop a general monitoring method applicable for both linear and nonlinear systems with multivariables. It also shows how static PCA can be used for the dynamic system. This technique, referred to as NNPCA, integrates NN with PCA. NN is employed to model the nonlinear dynamic system. The actual behavior of the process to be supervised is compared with that of a nominal fault-free neural network model driven by the same observations. The multivariable residuals derived from the differences between these outputs are evaluated by the PCA method. In other words, the proposed technique uses NN as the nonlinear dynamic operator to remove the nonlinear and dynamic characteristics and applies PCA to generating simple monitoring charts
The rest of this paper is organized as follows. A simple example demonstrates the different detectability between dynamic and static control charts for a dynamic system in Section 2. The dynamic process-monitoring scheme is proposed in Section 3, including the residual generator by dynamic neural networks and the residual evaluator by PCA. In Section 4, two case studies, a simulated Tennessee Eastman process and surface quality in a stainless steel slab, are employed to illustrate validity of the proposed technique. Finally, summaries and conclusions are presented.
Section snippets
Dynamic control charts
In traditional steady-state process monitoring, the Shewhart–Deming statistical model can be written asorwhere , the measurement of the process variable at time k, is represented by a fixed target mean μ plus a deviation from the target , often called the random measurement error resulted from the uncertain variations and disturbances among the lurking variables. The target mean μ is the function of x1,x2,⋯,xm that keeps the mean constant. When the process
Dynamic process monitoring scheme
The process monitoring structure for dynamic systems consists essentially of two core stages: residual generation and residual evaluation (Fig. 2). Residual generation is related to the actual behavior of the process to be supervised compared with that of nominal model-observation features driven by the same inputs. This allows finding a difference with respect to the normal operating condition. It is expected that the residual variables between the estimated value of the model and the actual
Industrial problem applications
The use of the neural network and PCA for dynamic statistical process monitoring is demonstrated through a complex Tennessee Eastman simulation process as well as a real industrial steel slab process for detecting the surface quality.
Conclusion
Process detection and diagnosis is currently one of the largest application domains of neural network systems. Strategies and capabilities for fault monitoring and diagnosis have been evolving rapidly. Most of the past applications involving monitoring and diagnosis were based on prediction residuals. That is, they used simple prediction errors for each variable to provide mapping between the possible causes and the possible faults. This approach is valid, however, only when the prediction
Acknowledgements
Support from China Steel Corporation is gratefully acknowledged. We are indebted to Dr. Muh-Jung Lu for giving us access to the steel slab data.
References (37)
Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy — a survey and some new results
Automatica
(1990)Process fault detection based on modeling and estimation methodsa survey
Automatica
(1984)Fuzzy systems design based on a hybrid neural structure and application to fault diagnosis of technical processes
Control Engineering Practice
(1996)- et al.
Fault diagnosis and neural network worksa power plant application
Control Engineering Practice
(1995) - et al.
Process fault detection and diagnosis using neural networks — I. Steady-state processes
Compu. Chem. Eng.
(1990) - et al.
The process chemometrics approach to process monitoring and fault detection
J. Proc. Cont.
(1996) Autoassociative neural networks
Comput. Chem. Eng.
(1992)- et al.
Nonlinaer principal component analysis — based on principal curves and neural network
Comput. Chem. Eng.
(1996) Exponentially weighted moving principal components analysis projections to latent structures
Chemometrics Intell. Lab. Syst.
(1994)- et al.
Disturbance detection and isolation by dynamic principal component analysis
Chemometrics Intell. Lab. Syst.
(1995)
Statistical process control of multivariate processes
Control Engineering Practice
A plant-wide industrial process control problem
Comput. Chem. Eng.
Base control for the Tennessee Eastman problem
Comput. Chem. Eng.
Fault Detection and Diagnosis in Chemical and Petrochemical Processes
Neural network in process fault diagnosis
IEEE Trans. Syst. Man Cyber.
Cited by (151)
Smart batch process: The evolution from 1D and 2D to new 3D perspectives in the era of Big Data
2023, Journal of Process ControlLong–short-term memory encoder–decoder with regularized hidden dynamics for fault detection in industrial processes
2023, Journal of Process ControlFault detection and diagnosis for non-linear processes empowered by dynamic neural networks
2022, Computers and Chemical EngineeringCitation Excerpt :One of the first attempts was from Lin et al. (2000) who used a nonlinear dynamic PCA with feed-forward neural network. Akbaryan and Bishnoi (2001) used Decision Trees, Kano et al. (2002) used moving principal component analysis, Chen and Liao (2002) used shallow Neural Networks with PCA, Maurya et al. (2005) used a PCA - Qualitative trend analysis, while (Chiang et al., 2004) and (Zhang, 2009) used Support Vector Machines (SVMs) and some alterations of them for fault classification. Also, a different approach from (Odiowei and Cao, 2010) with a novel state-space independent component analysis, Eslamloueyan (2011) used Hierarchical neural network,(Lau et al., 2013) used multi-scale PCA with adaptive neuro-fuzzy inference system, Rad and Yazdanpanah (2015) used multilayer perceptron based on Expectation-Maximization (EM) clustering.
Manifold regularized stacked autoencoders-based feature learning for fault detection in industrial processes
2020, Journal of Process Control