- Email: [email protected]
Assessment and Identification of Undesired States in Chemical Semibatch Reactors Using Neural Networks
Copyright @ IF AC Fault Detection, Supervision and Safety for Technical Processes, Budapest, Hungary. 2000
ASSESSMENT AND IDENTIFICATION OF UNDESIRED STATES IN CHEMICAL SEMIBATCH REACTORS USING NEURAL NETWORKS
G. Hessel, H. Kryk, W. Schmitt, T. Seiler, F.-P. Weiss, G. Decrbcrg*, J. Neumann*
Forschungszentrum Rossendorle. V, Jnstitut fur Sicherh eitsforschung Postjach 510119, D-O 1314 Dresden, Germany * Fraunhofer Institutjur Umwelt-, Sicherheits- und Energietechnik UMSICHT Osterfelder Str. 3, D-46047 Oberhausen, Germany E-mail : w.schmitl@fz-rossendorfde, Phone: (+49)351 260 3197, Fax (+49)351 2603651
Abstract: This paper presents a neural-network approach to operator-independent assessing the operational states of chemical semibatch reactors . The suitability of neural networks for process monitoring was investigated in a miniplant in which strongly exothermic chemical reference processes were carried out. Before being applied to state classification, the neural-network classifiers first have be trained using process data of normal and abnormal sequences of reaction to establish a nonlinear decision model between process parameters and state classification. Afterwards, the trained classifiers can be used for process monitoring. Best results were reached with three-layer perceptron networks . For assessing the danger potential of fault states, separate perceptron networks for danger classification and for fault identification were used. Copyright 2000 rf) IFAC Keywords : Fault diagnosis; Process identification; Supervision; Artificial intelligence; Classifiers; Neural networks; Chemical industry
I . INTRODUCTION
and to support them in decision making under critical situations.
Accident analyses have shown that human errors are one main cause of incidents with danger potential in chemical plants (Benuzzi and Zaldivar, 1991). The reasons for human errors under critical situations have mainly been the large amount of data which is generated in complex chemical plants and the overstraining of personnel due to an excessive presentation of information. Consequently, even experienced operators have difficulty to distinguish normal from abnormal operational conditions or to identify the cause of process trends. Therefore, the objective is to apply neural networks as an additional supervision system to diminish the load and stress on operators
In recent years, neural networks have been widely used in the field of chemical engineering, in process control, fault detection, and fault diagnosis. The neural network approach to fault diagnosis needs measurement values and the identified fault mode for each faulty operational state (Neumann, et aI. , 1998). However, data from all fault states cannot be obtained from chemical production plants in which strongly exothermic processes are carried out. For this purpose, these data have to be generated by means of a dynamic process simulator or/and by fault experiments carried out in a laboratory reactor capa-
447
ble of scaling up to the production plant. Since the development of a rigorous analytic modelling for nonlinear complex chemical processes is very timeand cost-intensive, this work aims at the use of training data generated in suited laboratory reactors.
relating the different features to the normal process state . For example, temperatures were normalized as a relative deviation from their normal set-point. Furthermore, all features of the input vector were normalized into the range from 0 to I. Beside extracting suitable features, it was also important to choose the appropriate type and the optimum topology of the neural network .
2. CLASSIFICATION TASK AND DATA The detection of a beginning incident in an early stage is an essential requirement for preventing accidents in chemical plants. Only if the causes of potentially dangerous situations arc recognized by the personnel in due time, it is possible to take the necessary counter measures to avoid a thermal explosion which might result in the release of the reaction mixture or even in the destruction of the plant. Since a short reaction time of the personnel, particularly, during transient operating conditions is of crucial importance, the efficiency of an additional supervi sion method based on pattern recognition and especially on neural networks should be proven in exothermic chemical reactions. As a reference process, the homogeneous catalytic esterification between acetic anhydride and methanol was chosen. This strongly exothermic reaction can be controlled within a certain temperature range. However, it has also enough self-reinforcing potential for heat production to investigate runaway phenomena in dependence on the temperature and on the concentration of the catalyst (sulphuric acid).
Fig . I. Identification of undesired operational states in a chemical plant
3. CLASSIFICATION OF OPERATIONAL STATES Since operating errors in the starting-up phase and in other transicnt conditions of batch and semi batch processes are a frequent cause of accidents, it is important to supervise the operational states to recognize human operational errors in an early stage. To investigate the classification of operational states, data of a semibatch process in the laboratory reactor were used rcpresenting several state classes like tempering (1), heating (2), calibration (3), cooling (4), dosing/reaction (5), danger of runaway (6) and slight dec rease of temperature (7).
Before applying to the qualitative classification of operational states, such a supervision method first has to be trained by the process data from normal operating conditions and different fault states as well. The scheme in Fig. I displays important components of the laboratory reactor and shows the working principle of a supervision method based on a neural network. To investigate the effect of different fault states on the reference process, typical faults were simulated in the laboratory reactor, e.g. loss of cooling, stirrer failure, inappropriate feed rate of reactants, inappropriate concentration of catalyst, different reaction temperatures, etc. These data sets were used for training and testing the state classifiers. In the training phase, the neural network is simultaneously confronted with pairs of feature vectors and target vectors at the input and output layer, respectively. To achieve a good classification ability, the input vectors have to consist of significant features characterizing the respective fault state, while the output vectors contain the expertise on the specific fault. The feature extraction for the input vector is based on both process signals (filling height, feed rates, stirrer speed, pressure, temperatures of reactor, jacket and of the coolant, etc .) and on calculated values from the measuring/controlling-computer of the laboratory reactor (e.g. thermal capacity and volume of the reaction mass, heat of reaction and reactor temperature gradient). To avoid multiple neural networks each for a special set-point of temperature or pressure, all features are normalized by
Tme(mnJ
Fig. 2. State classification of a process sequence in the laboratory reactor using the perceptron network (19/1517) Figure 2 illustrates the state classi fication of a threelayer perceptron network (1911517) with 19, 15 and 7 neurons in the input, hidden and output layer, respectively. Its seven output neurons correspond to the
448
several operating state classes characterized by differen! numbers, while the thickness of the layer is IdentIcal to the class membership in percent. The class membership corresponds to the value of the respective output neuron between 0 and I during the classification of pattern vectors. At the bottom of Fig. 2, the experts classification of the process states which was used for training is depicted in a reduced diagram. Contrary to the experts classification, the neural network already assigns the state "cooling" (4) and the state "tempering" (I) before the end of dosing (t = 155 min). The reason for this is that the reactant of the hold-up is completely consumed. Therefore, there is no hcat production but there arc still heat losses. In general, there exist no important differences between the experts decision and the network classification (Fig. 2).
with high class membcrship, while the normal vectors between 3923 and 3940 were assigned to the prealarm stage. However, that can be accepted because these reference vectors characterize a begin of dosing dUrIng a normal operational state. After having detected a deviation from the approved process control. a second neural network for fault isolation is used.
'00'
200.
300'
.()().
'O'"
Number 01 le.. hJ"" vector
4. DANGER CLASSIFICATION AND FAULT ISOLATION
Fig.3. Reclassification of data used to train the perceptron network (15/4/5) for danger classification
A neural network trained for classification of operational states is capable of correctly classifying normal states and several fault states, but it cannot provide InformatIOn on the potential danger of the different faults (Hessel, et al., 1997). Since the assessment of the danger and the early warning of the personnel arc important for taking suitable counter measures in dangerous states, a new supervision concept was realtzed. For this purpose, separate neural networks are used for classifying the danger and for fault isolation, respectively. As it can be seen in Fig. 3 and 4, five warning stages were chosen: normal (I), prealarm (2), alarm (3), extreme danger (4) and decreasing danger (5). To get a good generalization capability, the training data were com-posed of feature vectors from state changes of the defined warning stages which were chosen by the expert from all the available fault simulation experiments. For improving the sensitivity regarding a beginning dangerous fault, the features were mainly extracted from dosing rates, values of the pressure and the reactor temperature and their derivates, the second derivate of the reactor temperature, the temperature difference between reactor and jacket and its rate, the relative deviation of the reactor temperature from the temperature set-point and several values of temperature differences of the cooling system which were normalized regarding the temperature set-point. In total, there are 15 elements in the feature vector.
The following six typical fault classes were trained: stirrer failure, loss of cooling, low and high dosing rate, extraneous substance in the reaction mixture (mischarging) and false temperature level (temperature fault) . The training data are based on the same simulation experiments as used for danger classificatIOn, but the data ranges were newly chosen for the fault reference vectors. The reference vectors consisted of the same features. The perceptron network (15/3/6) trained for fault isolation can reclassify these training data very well. Even untrained data of two simultaneously occurring faults are unambiguously classified as shown at the bottom of Fig. 4. The unknown data come from a semi batch process where there are a high dosing rate a~d a stirrer failure simultaneously. At the upper part of FIg. 4, the profiles of some process and plant variables are presented, such as the reactor temperature Tr, the jacket temperature Tj, the temperature of the coolant reservoir Tc, the stirrer speed N, and the mass of the second reactant M added to the reaction mixture. I? the middle part, the results of the danger claSSIficatIOn are depicted for comparison. The fault network recognizes the stirrer failure and the high dOSIng rate where the maximum class membership of the hIgh dOSIng rate during the stirrer failure is not reached (see at the bottom of Fig. 4). Obviously, this neural network takes into consideration that the heat production due to the loss of mixing is lower than during normal operation. After restarting the stirrer, the rate of reaction increases exponentially due to the accumulated mass of reactants and due to the insufficient cooling. This is indicated by the short occurrence of the fault - loss of cooling. After that only the fault "high dosing rate" is classified because the exothermic reaction is restabilized by more intensive cooling of the jacket.
Figure 3 shows the reclassification results of the perceptron network ( 15/4/5) trained for danger classification. To illustrate the simultaneous occurrence of multiple faults, the summed class membership was introduced where each output neuron (fault) can be active with a maximum value of I. Although a summed class membership was used, more than 99% of the 5686 reference vectors were unambiguously assIgned regarding the desired warning stages. In Fig. 3, only the feature vectors between number 2610 and 2647 defined as an alarm stage were not recognized
449
provide sufficient time for the personnel to correct the deviation from safe operation. Current work is dealing with the application of this neural-network approach to monitor a heterogeneous hydrogenation process in a production plant.
The danger classifier provides reasonable information on the danger potential in this incident scenario. After the stirrer failure, first the neural network indicates a pre-alarm because of the missing dosing rate of the reactant. Then there is an alarm due to dosing of the second reactant. After restarting the stirrer, this extreme danger is correctly classified because, due to the reactant accumulation, a runaway reaction begins. However, the runaway reaction is stopped by an increased power of cooling. This is shown by thc classes "decreasing danger" and "pre-alarm" in the following time. It can be stated that the untrained incident scenario is correctly assigned and assessed by both perceptron networks.
6. ACKNOWLEDGMENTS This work is funded by the BMBF (Bundesministerium fUr Bildung, Wissenschaft, Forschung und Technologie) in Germany, Grant No. 01RG94235.
7. REFERENCES Benuzzi, A. and Zaldivar, 1. M. (eds.) (1991). Safety of Chemical Batch Reactors and Storage Tanks. Kluwer Academic Publishers, Dordrecht Neumann, 1.; Schluter, S.; Weinspach, P.-M.; Hessel, G.; Schmitt, W.; van der Vorst, K.; Wei13, F.-P.; (1998), Frilherkennung sicherheitsrelevanter Betriebszustande in Chemieanlagen mil neuronalen Netzen. at (Automatisierungstechnik) Heft 2, S. 104-110 Hessel, G.; Schmitt, W.; van der Vorst, K.; Wei13, F.-P.; Neumann, 1.; Schli.iter, S.; (1997). Identification of Dangerous States in Chemical Batch Reactors Using Neural Networks. Proc. of IF AC Symposium on Fault Detection, Supervision and Safety for Technical Processes SAFEPROCESS'97, Hull, UK, pp. 926-931
T..... [nc!
2.S
~
2
~
1.5
~ ~ 1
..'ii
E ~ 0 .5
1500
2100
2700
3300
3900
Time IsecJ
: I
Temperature fautJ
"
~
~ E
~
le
u
1
0
MIscharging
i HIgh dosing rate i
i
I
Low dosing rate
Loss of cooling i Stirrer faIlure
1500
.........-!-4r-
I.
i
I 2100
I 2700
3300
3900
TIme Isec]
Fig.4. Profiles of process variables and results of perceptron networks for danger classification and fault isolation during an incident scenario with stirrer failure and high dosing rate .
5. CONCLUSIONS The efficiency of the neural-network approach to assess and identify undesired operational states could be proven by data sets of a homogeneous esterification process. Results show that the multi-layer perceptron network is capable of correctly classifying normal operational states and faults of the investigated process sequences. However, a neural network trained for classification of operational states cannot assess the potential danger of different faults . Therefore, separate perceptron networks were used for danger assessment and for fault identification. Early warning and identification the fault causes should
450
ASSESSMENT AND IDENTIFICATION OF UNDESIRED STATES IN CHEMICAL SEMIBATCH REACTORS USING NEURAL NETWORKS
G. Hessel, H. Kryk, W. Schmitt, T. Seiler, F.-P. Weiss, G. Decrbcrg*, J. Neumann*
Forschungszentrum Rossendorle. V, Jnstitut fur Sicherh eitsforschung Postjach 510119, D-O 1314 Dresden, Germany * Fraunhofer Institutjur Umwelt-, Sicherheits- und Energietechnik UMSICHT Osterfelder Str. 3, D-46047 Oberhausen, Germany E-mail : w.schmitl@fz-rossendorfde, Phone: (+49)351 260 3197, Fax (+49)351 2603651
Abstract: This paper presents a neural-network approach to operator-independent assessing the operational states of chemical semibatch reactors . The suitability of neural networks for process monitoring was investigated in a miniplant in which strongly exothermic chemical reference processes were carried out. Before being applied to state classification, the neural-network classifiers first have be trained using process data of normal and abnormal sequences of reaction to establish a nonlinear decision model between process parameters and state classification. Afterwards, the trained classifiers can be used for process monitoring. Best results were reached with three-layer perceptron networks . For assessing the danger potential of fault states, separate perceptron networks for danger classification and for fault identification were used. Copyright 2000 rf) IFAC Keywords : Fault diagnosis; Process identification; Supervision; Artificial intelligence; Classifiers; Neural networks; Chemical industry
I . INTRODUCTION
and to support them in decision making under critical situations.
Accident analyses have shown that human errors are one main cause of incidents with danger potential in chemical plants (Benuzzi and Zaldivar, 1991). The reasons for human errors under critical situations have mainly been the large amount of data which is generated in complex chemical plants and the overstraining of personnel due to an excessive presentation of information. Consequently, even experienced operators have difficulty to distinguish normal from abnormal operational conditions or to identify the cause of process trends. Therefore, the objective is to apply neural networks as an additional supervision system to diminish the load and stress on operators
In recent years, neural networks have been widely used in the field of chemical engineering, in process control, fault detection, and fault diagnosis. The neural network approach to fault diagnosis needs measurement values and the identified fault mode for each faulty operational state (Neumann, et aI. , 1998). However, data from all fault states cannot be obtained from chemical production plants in which strongly exothermic processes are carried out. For this purpose, these data have to be generated by means of a dynamic process simulator or/and by fault experiments carried out in a laboratory reactor capa-
447
ble of scaling up to the production plant. Since the development of a rigorous analytic modelling for nonlinear complex chemical processes is very timeand cost-intensive, this work aims at the use of training data generated in suited laboratory reactors.
relating the different features to the normal process state . For example, temperatures were normalized as a relative deviation from their normal set-point. Furthermore, all features of the input vector were normalized into the range from 0 to I. Beside extracting suitable features, it was also important to choose the appropriate type and the optimum topology of the neural network .
2. CLASSIFICATION TASK AND DATA The detection of a beginning incident in an early stage is an essential requirement for preventing accidents in chemical plants. Only if the causes of potentially dangerous situations arc recognized by the personnel in due time, it is possible to take the necessary counter measures to avoid a thermal explosion which might result in the release of the reaction mixture or even in the destruction of the plant. Since a short reaction time of the personnel, particularly, during transient operating conditions is of crucial importance, the efficiency of an additional supervi sion method based on pattern recognition and especially on neural networks should be proven in exothermic chemical reactions. As a reference process, the homogeneous catalytic esterification between acetic anhydride and methanol was chosen. This strongly exothermic reaction can be controlled within a certain temperature range. However, it has also enough self-reinforcing potential for heat production to investigate runaway phenomena in dependence on the temperature and on the concentration of the catalyst (sulphuric acid).
Fig . I. Identification of undesired operational states in a chemical plant
3. CLASSIFICATION OF OPERATIONAL STATES Since operating errors in the starting-up phase and in other transicnt conditions of batch and semi batch processes are a frequent cause of accidents, it is important to supervise the operational states to recognize human operational errors in an early stage. To investigate the classification of operational states, data of a semibatch process in the laboratory reactor were used rcpresenting several state classes like tempering (1), heating (2), calibration (3), cooling (4), dosing/reaction (5), danger of runaway (6) and slight dec rease of temperature (7).
Before applying to the qualitative classification of operational states, such a supervision method first has to be trained by the process data from normal operating conditions and different fault states as well. The scheme in Fig. I displays important components of the laboratory reactor and shows the working principle of a supervision method based on a neural network. To investigate the effect of different fault states on the reference process, typical faults were simulated in the laboratory reactor, e.g. loss of cooling, stirrer failure, inappropriate feed rate of reactants, inappropriate concentration of catalyst, different reaction temperatures, etc. These data sets were used for training and testing the state classifiers. In the training phase, the neural network is simultaneously confronted with pairs of feature vectors and target vectors at the input and output layer, respectively. To achieve a good classification ability, the input vectors have to consist of significant features characterizing the respective fault state, while the output vectors contain the expertise on the specific fault. The feature extraction for the input vector is based on both process signals (filling height, feed rates, stirrer speed, pressure, temperatures of reactor, jacket and of the coolant, etc .) and on calculated values from the measuring/controlling-computer of the laboratory reactor (e.g. thermal capacity and volume of the reaction mass, heat of reaction and reactor temperature gradient). To avoid multiple neural networks each for a special set-point of temperature or pressure, all features are normalized by
Tme(mnJ
Fig. 2. State classification of a process sequence in the laboratory reactor using the perceptron network (19/1517) Figure 2 illustrates the state classi fication of a threelayer perceptron network (1911517) with 19, 15 and 7 neurons in the input, hidden and output layer, respectively. Its seven output neurons correspond to the
448
several operating state classes characterized by differen! numbers, while the thickness of the layer is IdentIcal to the class membership in percent. The class membership corresponds to the value of the respective output neuron between 0 and I during the classification of pattern vectors. At the bottom of Fig. 2, the experts classification of the process states which was used for training is depicted in a reduced diagram. Contrary to the experts classification, the neural network already assigns the state "cooling" (4) and the state "tempering" (I) before the end of dosing (t = 155 min). The reason for this is that the reactant of the hold-up is completely consumed. Therefore, there is no hcat production but there arc still heat losses. In general, there exist no important differences between the experts decision and the network classification (Fig. 2).
with high class membcrship, while the normal vectors between 3923 and 3940 were assigned to the prealarm stage. However, that can be accepted because these reference vectors characterize a begin of dosing dUrIng a normal operational state. After having detected a deviation from the approved process control. a second neural network for fault isolation is used.
'00'
200.
300'
.()().
'O'"
Number 01 le.. hJ"" vector
4. DANGER CLASSIFICATION AND FAULT ISOLATION
Fig.3. Reclassification of data used to train the perceptron network (15/4/5) for danger classification
A neural network trained for classification of operational states is capable of correctly classifying normal states and several fault states, but it cannot provide InformatIOn on the potential danger of the different faults (Hessel, et al., 1997). Since the assessment of the danger and the early warning of the personnel arc important for taking suitable counter measures in dangerous states, a new supervision concept was realtzed. For this purpose, separate neural networks are used for classifying the danger and for fault isolation, respectively. As it can be seen in Fig. 3 and 4, five warning stages were chosen: normal (I), prealarm (2), alarm (3), extreme danger (4) and decreasing danger (5). To get a good generalization capability, the training data were com-posed of feature vectors from state changes of the defined warning stages which were chosen by the expert from all the available fault simulation experiments. For improving the sensitivity regarding a beginning dangerous fault, the features were mainly extracted from dosing rates, values of the pressure and the reactor temperature and their derivates, the second derivate of the reactor temperature, the temperature difference between reactor and jacket and its rate, the relative deviation of the reactor temperature from the temperature set-point and several values of temperature differences of the cooling system which were normalized regarding the temperature set-point. In total, there are 15 elements in the feature vector.
The following six typical fault classes were trained: stirrer failure, loss of cooling, low and high dosing rate, extraneous substance in the reaction mixture (mischarging) and false temperature level (temperature fault) . The training data are based on the same simulation experiments as used for danger classificatIOn, but the data ranges were newly chosen for the fault reference vectors. The reference vectors consisted of the same features. The perceptron network (15/3/6) trained for fault isolation can reclassify these training data very well. Even untrained data of two simultaneously occurring faults are unambiguously classified as shown at the bottom of Fig. 4. The unknown data come from a semi batch process where there are a high dosing rate a~d a stirrer failure simultaneously. At the upper part of FIg. 4, the profiles of some process and plant variables are presented, such as the reactor temperature Tr, the jacket temperature Tj, the temperature of the coolant reservoir Tc, the stirrer speed N, and the mass of the second reactant M added to the reaction mixture. I? the middle part, the results of the danger claSSIficatIOn are depicted for comparison. The fault network recognizes the stirrer failure and the high dOSIng rate where the maximum class membership of the hIgh dOSIng rate during the stirrer failure is not reached (see at the bottom of Fig. 4). Obviously, this neural network takes into consideration that the heat production due to the loss of mixing is lower than during normal operation. After restarting the stirrer, the rate of reaction increases exponentially due to the accumulated mass of reactants and due to the insufficient cooling. This is indicated by the short occurrence of the fault - loss of cooling. After that only the fault "high dosing rate" is classified because the exothermic reaction is restabilized by more intensive cooling of the jacket.
Figure 3 shows the reclassification results of the perceptron network ( 15/4/5) trained for danger classification. To illustrate the simultaneous occurrence of multiple faults, the summed class membership was introduced where each output neuron (fault) can be active with a maximum value of I. Although a summed class membership was used, more than 99% of the 5686 reference vectors were unambiguously assIgned regarding the desired warning stages. In Fig. 3, only the feature vectors between number 2610 and 2647 defined as an alarm stage were not recognized
449
provide sufficient time for the personnel to correct the deviation from safe operation. Current work is dealing with the application of this neural-network approach to monitor a heterogeneous hydrogenation process in a production plant.
The danger classifier provides reasonable information on the danger potential in this incident scenario. After the stirrer failure, first the neural network indicates a pre-alarm because of the missing dosing rate of the reactant. Then there is an alarm due to dosing of the second reactant. After restarting the stirrer, this extreme danger is correctly classified because, due to the reactant accumulation, a runaway reaction begins. However, the runaway reaction is stopped by an increased power of cooling. This is shown by thc classes "decreasing danger" and "pre-alarm" in the following time. It can be stated that the untrained incident scenario is correctly assigned and assessed by both perceptron networks.
6. ACKNOWLEDGMENTS This work is funded by the BMBF (Bundesministerium fUr Bildung, Wissenschaft, Forschung und Technologie) in Germany, Grant No. 01RG94235.
7. REFERENCES Benuzzi, A. and Zaldivar, 1. M. (eds.) (1991). Safety of Chemical Batch Reactors and Storage Tanks. Kluwer Academic Publishers, Dordrecht Neumann, 1.; Schluter, S.; Weinspach, P.-M.; Hessel, G.; Schmitt, W.; van der Vorst, K.; Wei13, F.-P.; (1998), Frilherkennung sicherheitsrelevanter Betriebszustande in Chemieanlagen mil neuronalen Netzen. at (Automatisierungstechnik) Heft 2, S. 104-110 Hessel, G.; Schmitt, W.; van der Vorst, K.; Wei13, F.-P.; Neumann, 1.; Schli.iter, S.; (1997). Identification of Dangerous States in Chemical Batch Reactors Using Neural Networks. Proc. of IF AC Symposium on Fault Detection, Supervision and Safety for Technical Processes SAFEPROCESS'97, Hull, UK, pp. 926-931
T..... [nc!
2.S
~
2
~
1.5
~ ~ 1
..'ii
E ~ 0 .5
1500
2100
2700
3300
3900
Time IsecJ
: I
Temperature fautJ
"
~
~ E
~
le
u
1
0
MIscharging
i HIgh dosing rate i
i
I
Low dosing rate
Loss of cooling i Stirrer faIlure
1500
.........-!-4r-
I.
i
I 2100
I 2700
3300
3900
TIme Isec]
Fig.4. Profiles of process variables and results of perceptron networks for danger classification and fault isolation during an incident scenario with stirrer failure and high dosing rate .
5. CONCLUSIONS The efficiency of the neural-network approach to assess and identify undesired operational states could be proven by data sets of a homogeneous esterification process. Results show that the multi-layer perceptron network is capable of correctly classifying normal operational states and faults of the investigated process sequences. However, a neural network trained for classification of operational states cannot assess the potential danger of different faults . Therefore, separate perceptron networks were used for danger assessment and for fault identification. Early warning and identification the fault causes should
450