Contact a Yokogawa Expert to learn how we can help you solve your challenges.
Management in the Chemical Process Industry
Diego Ruiz,† Chouaib Benqlilou,† Jose ́ M. Nougue ́s,† and Luis Puigjaner*
Chemical Engineering Department, Univesitat Polite`cnica de Catalunya, Av. Diagonal 647, 08028 Barcelona, Spain
Advanced Control Group, SOTEICA SRL, Alvarez Thomas 796, 3 C, 1427 Buenos Aires, Argentina
The aim of this work is to present a proposal for implementation of a support framework for abnormal situation management in the chemical process industry. A main feature of the technology developed is that it takes advantage of existing software packages that are familiar to plant engineers (e.g. Plant Information System) and a commercial process simulator. On the basis of three sources of information (a historical database, a HAZOP analysis, and a first principles plant model), the support framework is developed and easily implemented into the real plant. It consists of a preprocessing module, which performs a variety of key tasks using plant data such as data reconciliation, filtering, and denoising. Some of the outputs of this preprocessing module are the inputs of the fault diagnosis system (FDS). This FDS is a combination of a pattern recognition approach based on neural networks and a fuzzy logic system (FLS) in a block oriented configuration. The case study used to demonstrate the FDS implementation corresponds to a real petrochemical plant.
Troubleshooting in chemical process plants is one area where computer technology can now provide substantial benefits. Computers are powerful enough to handle meaningful systems. It is now possible to deliver large scale knowledge-based systems directly to the end-user. Furthermore, current chemical plants save large amounts of data containing valuable information about the processes. In this context, the use of combinations of historical-based methods and knowledge-based techniques for fault diagnosis in chemical plants shows promising results.1 Research is focused on the simplicity of implementation and the optimal performance of such systems.
The problem of abnormal situation management is made considerably difficult by the scale and complexity of modern plants. Dash and Venkatasubramanian2 have reported that a hybrid blackboard-based framework utilizing collective problem solving is shown to be the most promising approach. However, its implementation does not seem very easy.
In this work, a proposal for implementation of a support framework for abnormal situation management in the chemical process industries is shown. On the basis of three sources of information (a historical database, a HAZOP analysis, and a regular plant model), the support framework is developed and easily implemented into the real plant. It consists of a preprocessing module which performs a variety of key tasks using plant data such as trend generation, principal component analysis, data reconciliation, filtering, and denoising. Some of the outputs of this preprocessing module are the inputs of the fault diagnosis system (FDS). This FDS is a combination of a pattern recognition approach based on artificial neural networks (ANNs) and a fuzzy logic system (FLS) in a block oriented configuration. The FLS also has the function of alarm handling. The outputs of the FDS can be used by an advanced control module in order to take control actions, or by the operators who are responsible for decisionmaking, or by other levels in the information system as the scheduling system.
The case study selected to show the FDS implementation corresponds to a real petrochemical plant consisting of a train of distillation columns where a group of n-paraphines are separated from kerosene. Fluctuations in pressure of the hot oil to the reboilers have been considered as suspected faults.
The paper is organized as follows. First, the implementation of the system based on the different sources of information is explained, including some details of the data reconciliation module. Then, the development of the fault diagnosis system within the support framework is summarized. Next, the actual petrochemical plant is briefly described. Finally, implementation results are shown and conclusions given.
The information needed to implement the FDS includes a historical database, a hazard and operability (HAZOP) analysis, and a model of the chemical plant. These sources of information will be used by the different components of the FDS, as is described next.
Presently, process computers collect a massive amount of data from a multitude of plant sensors every few minutes or seconds. Present chemical plants do not take appropriate advantage of this powerful source of information. A number of applications in process modeling, monitoring, and control can be actually performed.
The historical database that includes information related to normal and abnormal operating conditions can be used to train the ANN structure. The on-line measurements from the plant are the ANNs inputs. The ANNs outputs, the so-called "residuals" in fault diagnosis, are the signals of different suspected faults. These outputs are a subset of the set of FLS inputs. The ANN not only has the advantage of a classifier but also can be retrained during use. In this way, the changing operating conditions of chemical plants do not affect the FDS performance.
A model is used at different design levels of a plant. Nowadays, the use of a commercial process simulator allows us to develop plant models easily enough. This practice is successful mainly in the petrochemical industry. Process engineers handle different kinds of plant models, and the model does not have to be completely accurate. It has to be accurate enough to give a satisfactory performance in the system using it.
A model of the plant can be used to obtain plant operation experience through simulation. The simulation can provide data on infrequent faults because in the cases of faults that rarely occur it is not possible to test the FDS using only plant data. In addition, by testing the ANN with the model, the development of the rules on the basis of experience with ANN performance is straightforward. The model is also useful for testing and validating the FDS.
In most industrialized countries, standards require us to perform hazard analysis on a regular basis. HAZOP is the most widely used and is recognized as the preferred approach in the chemical process industry. It is typically performed by a team of experts having specialized knowledge and expertise in the design, operation, and maintenance of the plant. HAZOP is also one of the most powerful hazard identification methods available and has been well described in the literature. Wennersten et al.4 have considered the extension of the HAZOP method to fault diagnosis characterization, concluding that it provides a more "down to the earth" approach for implementing an operator support system. The rules are kept simple to avoid the general problems of large rule-based knowledge systems, such as contradictory rules, large amounts of irrelevant information, and complex tree structures.
In general, the term "Causes" in the HAZOP analysis corresponds to the root cause of a deviation and it can be considered with the term fault. Only the causes that have been defined in the set of faults are considered in this step.
The HAZOP analysis allows us to generate the IF-THEN rules for the fuzzy logic system (FLS) and to determine the information to be sent to other levels in the information system. Not all the variables considered in the HAZOP analysis are measurements from the plant. Some variables can be observed (estimated). Regarding the generation of IF-THEN rules, only the measured and the observable variables should be considered. In section 6.5 (Implementation Results), the development of the IF-THEN rules for the petrochemical plant from the HAZOP analysis is shown.
Reliable and accurate process measurements are the key to efficient operation of chemical plants. It should be common practice to condition process measurements, so that measurement noises or other types of errors that affect data accuracy are eliminated or compensated. In general, data inherently contain inaccurate information, since measurements are obtained with imperfect instruments.
Data reconciliation (DR) is a procedure that makes the raw measurements consistent with the conservation's laws and other exact constraints, so that the random errors are eliminated and the variables' variance is reduced. The quality of a measured data set can be improved, since there is some degree of redundancy among the measured variables. Indeed the aim of data reconciliation is to eliminate the conflict among the redundant variables. The data reconciliation module is an important element in any abnormal situation management system.
There are many works addressing various computational and accuracy aspects in steady-state and dynamic data reconciliation. Crowe5 reviews the issue of reconciling a process measurement in a widely used and illustrative study. In a recent book,6 a systematic and comprehensive treatment of data reconciliation is provided.
In this work the DR algorithm and the corresponding quadratic programming resolution have been implemented by using Matlab. Nevertheless, nowadays it is possible to find a set of commercial data reconciliation software (e.g. Vali III from Belsim, DATACON) that could give the same results with less time spent for the development and implementation.
A key element to manage abnormal situations in a plant is a robust FDS. As can be seen in Figure 1, the FDS receives plant measurements and is able to detect deviations from normal operating conditions and also can determine their root cause. A fast identification of a fault can be utilized by a scheduling system to update the schedule in the most effective way, by the control system in order to take automated control actions, and by the operators, as a support for decision-making. The main objective is to avoid plant shutdowns. The plant should continue working satisfactorily despite the faults. In this way the productivity can be substantially increased.
The proposed FDS consists of a combination of a pattern recognition approach and an inference system. With historical data taken from the historical database of past faults, an ANN is trained to classify them. On the other hand, from the HAZOP analysis already described, a set of IF-THEN rules are defined. This set is completed with those coming from the experience with the use of the ANN.
An important aspect for the successful implementation of the FDS is the feature extraction to generate the patterns used in the ANN training. The problem of the traditional ANNs related to totally capture the space and time characteristics of process signals is overcome with the use of wavelet functions. Plant signals are preprocessed by a multiscale wavelet decomposition; then the extrema of a high level of detail are determined and input to the ANN classifier.7 Figure 2 shows the information flow in the proposed FDS. The outputs of the ANN are inputs of the fuzzy logic inference system (FLS). The fault signal has a value between 0 (no fault) and 1, corresponding to a specified suspected fault.
The case study corresponds to a real petrochemical plant consisting of a train of two distillation columns where a group of n-paraphines are separated from kerosene. Figure 3 shows the plant flow sheet taken from the HYSYS‚plant simulation interface. The feed to the plant consists of a mix of hydrocarbons that is preheated in a heat exchanger which takes advantage of a lateral extraction of the second column (redistillation column). The light hydrocarbons (less than C-10) and sulfur are separated in the first column (stripper) at the top (light kerosene). The stripper's bottom is fed to the redistillation column. At the top, the main product containing lineal hydrocarbons (C-10 to C-14) is obtained. At the bottom heavy kerosene is obtained as a byproduct.
The interaction between the two columns is due to two facts: energy exchange between connecting flows and the linking of the stripper's bottom, which feeds the redistillation column.
Further details of the process are withheld for commercial confidentiality reasons.
Data were collected using a plant information (PI) historian Excel Data Link interface, collecting 50 operating variables during the period December 1998 to February 1999. After data screening, inspection to identify unstable operating conditions or sensor malfunction, a 5 day period was selected as a training set.
The NeurOnline Studio can be used off-line or on-line. It is a tool for analysis of processes. Typically, the source of data is a data historian or other data archive. Here, it has been used as a visualization tool for principal component analysis (PCA) plots. Figure 4 shows the plot of the two first principal components. Three points can be distinguished corresponding to an identified deviation (highlighted with white boxes).
Using the step-by-step guidance provided by NeurOnline Studio, an ANN has been trained and tested to predict the identified upset. Figure 5 shows the output of the ANN model, which gives the value of 1 when the fault occurs. As can be observed, the predicted values practically overwrite the actual ones.
To show a more general applicability of the proposed FDS, another kind of ANN, programmed with Matlab, was utilized in section 6.5 (Implementation Results)
In this work HYSYS.Plant, a commercial simulator from AEA Technology-Hyprotech, has been used to simulate the plant. It has been chosen as an example of how to use existing available software in the proposed FDS implementation strategy.
Table 1 shows a simplified sample of a plant HAZOP analysis which has as cause the considered fault "high hot oil to the reboiler". The set of variables considered are only the on-line measurements from the plant.
In this industrial application, a steady-state data reconciliation has been performed in order to increase the quality of the input/ output mass flow rate of the process. Nevertheless, the flow rate of the light kerosene is not a measured variable. Therefore, to close the mass balance, a soft sensor has been implemented to estimate the value of this variable. An analysis of cause-effect and the knowledge of the interaction among the measured variables allow selecting the set of input variables that can better predict the desired output variables.
First of all, the selected input variables were filtered and denoised using wavelets and then used to construct the empirical model. On the basis of the fact that an ANN model presents a good ability to represent highly nonlinear systems and handle well the multi-input single-output (MISO) process, it has been adopted as the empirical model to estimate the output variables. The measured variables obtained from the physical sensors and those obtained from the soft sensors form a redundant model. This model is used within the data reconciliation algorithm in order to improve the quality of the variable considered. This approach is presented in Figure 6.
A soft sensor could be a good alternative to physical sensors in two main aspects: maintenance and capital cost reduction and also in risk reduction of sensor failure. The drawback of a soft sensor (ANN model) in terms of extrapolation and prediction could be the same as that of the physical sensor. Therefore, both sensors are similar in this respect.
Figure 7 shows the raw measurement and the filtered data. The accumulation mass and reflux flow rate were chosen as inputs to the model while the distillate mass flow rate is the model output. The distillate mass flow
rate was used to train the empirical model, and it was obtained by an observable equation (mass balance). Following Gertler,8 the ANN used was a feed-forward back-propagation with a moving window where the current and past values of the level in the accumulator (Acum in the flow sheet, Figure 3) and reflux are fed with the past values of the distillate as ANN input; the output of the ANN is the current distillate value. The size of the windows and the number of neurons in the hidden layer were chosen by trial and error. The performance of the training and validation of the empirical model used as a soft sensor can be seen in Figure 8.
Figure 9 shows the reconciled variables and the current measured ones. It can be seen that the correction affects the physical sensor as well as the soft sensor, and that the measured variable is consistent with the total mass balance. It is important to mention that the light kerosene flow rate was obtained using a soft sensor based on other measured variables, as has been explained in the previous paragraphs. It is also worth mentioning that testing was carried out with data which were not considered during the construction of the empirical model.
The implementation test has been focused on a special problem which corresponds to a fluctuation in the pressure of hot oil utilized for the reboiler's heating at the redistillation column. This fault has been simulated, and the variable profiles were saved.
The different signals have been processed by a multilevel 1-D wavelet using a specific wavelet filter. An eight-coefficient Daubechies wavelet has been used as a filter. Because the noise components are reduced and then disappear as the scale increases, the detail of scale 5 was considered sufficient. Then, the extrema of the processed signals have been determined for each signal, following the methodology proposed by Chen et al.
The following measured process variables show special features corresponding to the studied fault: feed to the stripper flow rate; stage 35 temperature in the stripper column; boilup and stage 38 temperatures in the redistillation column; and reflux to redistillation column flow rate. These patterns have been used as inputs to the ANN classifier. Figure 10 shows the patterns corresponding to the signal of the extrema determination from wavelet decomposition of the signal feed flow rate to the stripper. It can be seen how the frequency of the extrema is increased with the size of the fault. The ANN architecture used as a classifier in this study was a probabilistic ANN, which is a kind of radial basis network suitable for classification problems.
The development of the rules has been carried out from the HAZOP analysis. The set of IF-THEN rules extracted from the HAZOP analysis shown in Table 1 is the following (the considered fault "high hot oil to the reboiler" is coded as FAULT 1):
IF LREB1 IS HIGH THEN FAULT1 IS HIGH
IF LREB3 IS LOW THEN FAULT1 IS HIGH
IF TTOP1 IS HIGH THEN FAULT1 IS HIGH
IF TTOP3 IS HIGH THEN FAULT1 IS HIGH
IF TBOT3 IS HIGH THEN FAULT1 IS HIGH
The previous set must be completed considering the opposite changes with the consequent "FAULT1 IS LOW" and the IF-THEN rules which include the ANN classifier information (e.g. "IF ANNFault1 IS HIGH THEN FAULT1 IS HIGH"). The FLS is of Sugeno type.
The real plant has been simulated with HYSYS.Plant, using the DCS driver to allow communication with other applications. In this case, the other application is MATLAB, where the preprocessing system and the FDS are run.
Figure 11 shows the graphical user interface of HYSYS.Plant with the strip charts of the measured variables and the fault signal in the top right corner. The fault signal jumps from zero to one.
The fault isolation is shown to be fast enough to take actions on to the plant.
A proposal to speed up the implementation of a support framework for abnormal situation management in a real chemical plant has been shown through an industrial application.
The main feature of the technology developed is that it takes advantage of existing software packages that are familiar to plant engineers (e.g. plant information system) and a commercial process simulator. The fundamentals of the employed techniques such as preprocessing with wavelets, ANN training, and fuzzy logic system development have been treated in previous papers. This work focused on the integration and implementation in a real case of the developed technology. Furthermore, an important task, the data reconciliation, has been shown in some detail. It has been highlighted that the use of soft sensors increases data reconciliation performance.
Integration of the developed platform with the existing hardware and software in our RTO facilities is underway. Future work includes the application of the presented technology in a sugar cane refinery.
Collaboration with Repsol-YPF (Petroqu ́ımica La Plata, Argentina) is appreciated. NeruOn line Studio (Gensym) has been provided by SOTEICA S.R.L. AEA Technology-Hyprotech has provided HYSYS.Plant and its DCS Link, which were very useful in order to test the FDS in on-line mode. Finally, financial support from the European Community is gratefully acknowledged: Chem Project is funded by the European Community under the Competitive and Sustainable Growth Programme of the Fifth RTD Framework Programme (1998-2002) under Contract G1RD-CT-2001-00466. D.R. is sponsored by a Generalitat de Catalunya, II Pla de Recerca, TDOC Grant.
Contact a Yokogawa Expert to learn how we can help you solve your challenges.