integration of artificial intelligence systems for nuclear power plant ...

INTEGRATION OF ARTIFICIAL INTELLIGENCE SYSTEMS FOR NUCLEAR POWER PLANT SURVEILLANCE AND DIAGNOSTICS Robert E. Uhrig and J. Wesley Hines Department of Nuclear Engineering, University of Tennessee, Knoxville, TN 37996-2300 and William R. Nelson Human Factors Group, Idaho National Engineering and Environmental Laboratory, Idaho Falls, ID

ABSTRACT: The objective of this program is to design, construct operate, test, and evaluate a prototype integrated monitoring and diagnostic system for a nuclear power plant. It is anticipated that this technology will have wide application to other complex systems (e.g., fossil power plants, chemical processing plants, and possibly air traffic control systems). Over the past decade, the University of Tennessee (UT) and others have carried out many projects utilizing various methodologies of artificial intelligence (expert systems, neural networks, fuzzy systems and genetic algorithms) to enhance the performance (safety, efficiency, reliability, and availability) of nuclear power plants. Investigations and studies have included a) instrumentation surveillance and calibration validation, b) inferential sensing to calibration of feedwater venturi flow during fouling, c) thermodynamic performance modeling with iterative improvement of plant heat rate, d) diagnosis of nuclear power plant transients, and e) increase in thermal power through monitoring of DNBR (Deviation from Nucleate Boiling Regime). To increase the likelihood of these individual systems being used in a nuclear power plant, they must be integrated into a single system that operates virtually autonomously, collecting, interpreting, and providing information to the operators in a simple and understandable format.

DESCRIPTION OF PROJECT I. OVERVIEW Purpose. The purpose of this project is to integrate into a single system a variety of artificial intelligence (AI) based systems (expert systems, neural networks, fuzzy systems, and genetic algorithms) that can provide plant specific information to the plant operators in an intelligent, simple, understandable and non-intrusive manner regarding the status of the nuclear power plant. The validity/feasibility of each of the constituent applications of AI has been demonstrated to some extent. The principal investigator (Uhrig and Tsoukalas, 1997) recently identified some forty-five applications of artificial intelligence (Al) technologies (primarily neural networks and fuzzy systems) in nuclear power plants. Each of these applications has been demonstrated to provide some benefits to a nuclear power plant or its operators in the form of enhanced performance (safety, efficiency, reliability and/or availability). A few such systems may be worthy of stand-alone status in that the benefits justify an independent system, particularly when they are used outside the control room, but these are not included here. However, artificial intelligence based instruments that perform monitoring and diagnostics of the operation of the reactor or plant itself are not likely to be installed in the control room unless they can be integrated into a single system that operates virtually autonomously while providing information to the operators in a simple and understandable format. This project involves the design, development, construction, and testing of such an integrated non-invasive

1

system based on technical, economic, cost-benefit, human factors and human-machine interface considerations. The human factors and human-machine interface portion of this program will be guided by personnel from Idaho National Engineering and Environmental Laboratory (INEEL). The integrated system, which is the final product of this program, is called a “prototype.” This name properly reflects that it is an immediate predecessor to a commercial product that is consistent with the technology development roles of UT and INEEL. Data for test and evaluation of the prototype system will be obtained from the safety parameter display system, the plant computer of an operating nuclear power plants and/or a full scope, high fidelity simulator Pentium-based personal computers with 21- or 29-inch monitors will be used to simulate a typical control room display system. MATLAB and SIMULINK (software provided by MathWorks Inc. of Natick, MA) will be used for modeling the integrated system. After debugging and testing, the MATLAB and SIMULINK programs will be compiled into C++ language for operation, testing, and evaluation. Utility personnel, including reactor operators, will be involved during the design, testing, and evaluation of the integrated system. II. REVIEW OF CANDIDATE ARTIFICIAL INTELLIGENCE TECHNOLOGIES A preliminary review of the artificial intelligence projects already shown to be feasible for operational monitoring and diagnosis in nuclear power plants indicates that the following AI-based methodologies for surveillance and diagnostics are prime candidates for inclusion in the proposed integrated system. The final selection will be made with the concurrence of the utility/industrial participants in this program. Modular design of the systems will allow additional functions to be subsequently added and/or removed, if desired. A. Instrumentation, Surveillance and Calibration Verification. Traditional approaches to instrument calibration at nuclear power plants, especially instruments inside containment, are expensive in terms of labor, money, and radiation exposure. These calibrations require that the instrument be taken out of service and be falsely-loaded to simulate actual in-service stimuli. While proper adjustment is vital to maintaining proper plant operation, a less invasive technique is desirable. When implementing performance based calibrations, the instruments are calibrated only when necessary. On-line monitoring systems for calibration will allow utilities to determine when recalibration is needed, thereby reducing the frequency of calibration and the efforts necessary to assure the instruments continue to be calibrated properly. Benefits include very significant industry wide cost savings (utility estimates are about $1000 per calibration in containment), less time for reactor startup, and easier compliance with Nuclear Regulatory Commission (NRC) requirements for extending calibration intervals to match extended fuel cycles. A unique approach to "instrument surveillance and calibration verification" (ISCV) has been developed and tested at UT with the support of the U. S. Department of Energy (DOE) using a five-layer autoassociative neural network (AANN), a network where the outputs are trained to equal the inputs (Hines, Uhrig and Wrest, 1996). Digitized data available from the Safety Parameter Display System computer of Florida Power Corporation's Crystal River #3 Nuclear Power Plant has been used to demonstrate the validity of this approach. When the inputs are somewhat correlated, each output can be trained to be dependent upon all the correlated inputs as well as the corresponding input. The relationships between the different variables (i.e., the model of the system) are embedded in the weights by the training process. By introducing a systematic noise pattern with magnitudes up to 10% on top of the input training signals, the AANN is made very robust (i.e., each output is virtually immune to unanticipated changes in the corresponding input variable).

2

Once the network has been trained with normal operational data using this robust method, it monitors the selected plant variables to detect sensor drift or failure by simply comparing the network inputs with the corresponding best estimate outputs from the AANN. Any differences, called residuals, are then tested for significance using a very sensitive statistical method known as the Sequential Probability Ratio Test (SPRT). This SPRT based method is optimal in the sense that a minimum number of samples is required to detect a fault existing in the signal. If there is sensor drift, the residual mean shifts, and the SPRT parameters (specified by the operator through false and missed alarm probabilities) initiate an alarm when a sensor fails. The AANN also provides a correct surrogate signal that can temporarily replace the sensor signal until it can be repaired. The AANN method of monitoring many variables not only indicates when there is a sensor failure, it clearly indicates the signal charnel in which the signal error has occurred. Input variables to an AANN generally consist of primary and secondary side parameters such as flows, pressures, and temperatures that included many parameters considered by the regulatory authority to be safety parameters that must be monitored closely. In the United States, the Nuclear Regulatory Commission prescribes which variables must be monitored in the plant’s “Technical Specifications.” Genetic algorithms can be used to assure optimal grouping of variables into each of several AANNs (with 20 to 30 inputs each). The network uses non-linear activation functions for the three “hidden” layers and a linear activation function for the output layer. This linear layer allows for a fast mathematical regression technique, specifically "singular value decomposition" (SVD), to solve for the weights of the output layer of neurons. This also forces the nonlinear representation of the plant to be stored in the weights for the middle three layers. This regression feature also allows almost instantaneous “tuning" of the AANN (manual or automatic) when plant conditions change. An analysis of the sensitivity for changes in each of the networks outputs due to a 5% perturbation in each of it's input indicates that the robust training has made each output dependent on all the correlated inputs to about the same degree, and the sum of all errors due to these perturbations has been reduced by a factor of about 200. Benefits of this module: a) Assurance that sensors are in calibration, b) Ability to detect intermittent failures and noisy sensors, c) Availability of a surrogate sensor reading if needed, d) Ability to identify which sensor has drifted, became noisy, or failed, and e) Ability to differentiate between process change and sensor failure. B. On-Line Thermodynamic Performance Modeling and Improvement. In the past few years, several systems for monitoring heat rate performance of power plants based on "first principles" of energy balance applied to the many subsystems of power plants have been developed and integrated into an overall system for heat rate determination. An expert system combined with thermodynamic modeling to provide a reference heat rate is used to advise operators on steps to be taken to improve plant the heat rate. A potential drawback of this approach is that it is usually dependent upon system models based on ideal conditions, and often involves empirical relationships, and approximations of the actual processes, and linearizations of nonlinear phenomena. In an earlier UT study sponsored by DOE, a nonlinear thermodynamic process model was obtained using a neural network trained on actual thermodynamic measurements from the Sequoyah Nuclear Power Plant over a one-year period (Guo and Uhrig, 1992). Hence, the model represented the thermodynamic process as it actually existed in the plant, and the dynamic range of the data covered the normal range of variables during a typical annual cycle. More recent work utilizes genetic algorithms and principal component analysis to identify the optimal grouping of input parameters to the neural network models.

3

A neural network model of the thermodynamics of a power plant can be used to determine the influence of changes in different variable upon the heat rate through the use of sensitivity coefficients, where the signs indicate the directions of change in the variable that will improve heat rate, and the magnitude indicates the relative importance of the different variables. This information can be used to provide guidance to the plant operators and engineers as to where they should expend their efforts to improve the heat rate. An on-line heat rate monitor based on a neural network model (which is more related to actual conditions in a plant than models based on "first principles") can be utilized to determine which variables are the most important ones to adjust and whether they should be increased or decreased. If one or more of these variables are adjusted, the resultant change in heat rate can he monitored with the neural network model. Then another sensitivity measurement can be performed to indicate the next variable or set of variables that should be adjusted. This process can be continued on an iterative basis to achieve optimal efficiency under all existing or changing conditions (changing load, fouling of heat transfer surfaces, removal of components from service, changing air or river water temperature, etc.). Benefits of this system: a) Ability to iteratively move towards an optimal efficiency for existing conditions, b) A gain of $850,000 income per year for each 0.1% increase in efficiency, and c) Guidance to engineers as to how other improvements in thermodynamic efficiency can be obtained. C. Inferential Sensing and Virtual Measurements. Inferential sensing is often used to measure variables that cannot be measured directly. An inferential sensing system is defined as an instrumentation system which infers values of complex process variables by integrating information from multiple sensors. For instance, reactivity in nuclear power plants must be calculated from measurements and reactor design parameters. Neural networks can be trained to map almost instantaneously appropriate input variables into the desired output, e.g., reactivity. Inferential sensors, which incorporate a neural network for process modeling, can provide estimates of process variables that are usually measured off-line or through analytical laboratory instruments (e.g., such as the chemical composition of fluids). Work carried out by Kavaklioglu and Upadhyaya (1994) at UT included the inferential measurement of the flow of feedwater to the steam generators in a nuclear power plant after the feedwater venturi had been cleaned and calibrated. The neural network was then trained on data obtained during start up over the normal flow range, and put into a monitoring mode to predict the flow rate. The value predicted by the neural network remained relatively constant while the flow calculated from the venturi measurement increased due to fouling, approaching a near-asymptotic value about 0.8% higher in about one week and somewhat larger values later. Since the thermal power calibration of a nuclear power plant is directly dependent on the steam generator feedwater flow, an erroneous high reading gives a calculated power level that is higher than actual. Since nuclear power plants are usually licensed to a limiting thermal power rating, an erroneously high power level measurement effectively derates the plant. A fuzzy-neural methodology for monitoring unmeasurable variables in a complex system has also been developed at UT (Ikonomopoulos, Tsoukalas, and Uhrig, 1993). Neural networks map spatiotemporal information (in the form of time series) to algebraically defined fuzzy system membership functions. This entire process can be thought of as a virtual measurement. Through such virtual measurements, the values of (unmeasurable) monitored variables with operational significance (e.g., performance, valve position, or availability) can be evaluated. This methodology was applied to the High Flux Isotopes Reactor at Oak Ridge National Laboratory to evaluate the coolant control valve position. Benefits of this system: a) Ability to infer difficult to measure parameters, b) Avoidance of

4

derating due to feedwater venturi fouling (a 1% derating in a 1000 Mwe plant is worth about $2,500,000 per year) c) Ability to infer parameters that are normally calculated (e.g., reactivity) from measurements of related quantities, d) Ability to infer intangible parameters (e.g., quality, performance, etc.), and e) Ability to detect a deviation of performance from normal (e.g., fouling of heat transfer surfaces or pump degradation). D. Nuclear Power Plant Transients and Faults. When a nuclear power plant is operating normally, the readings of the instruments in a typical control room form a pattern (or unique set) of readings that represents a normal state of the plant or system. When a disturbance occurs, the instrument readings undergo a transition to a different pattern, representing a different state that may be normal or abnormal, depending upon the nature of the disturbance. The fact that the pattern of instrument readings undergoes a transition to a new state is sufficient to identify the fault or cause of the transient. Two different methods of identifying transients using neural networks have been explored at UT. Earlier work involved monitoring a small set of variables during the lifetime of the transients, including a reactor trip if one occurred. (Uhrig, 1989) The time records of these variables were digitized and applied to a neural network that had been trained using data from a simulation model or a full scope, high-fidelity plant simulator. The result is the classification of the transient or fault in accordance with the patterns of the transients utilized in the training of the neural network. More recent work by Bartlett and Uhrig (1992) and Guo and Uhrig (1992) involved the use of a large number (usually 20 to 40) of output variables from the plant that are sampled simultaneously, normalized to expected maximum values, preprocessed as necessary, and transmitted to the neural network. The unique relation among these 20 to 40 variables represents the condition of the plant at that particular instant. At a time ∆t after a transient begins, this relationship is very different, and it continues to change as the transient progresses. When a set of these sampled values are introduced into a trained neural network, it identifies the cause of the transient. Successive sets of sampled inputs (although different) will indicate that the same transient is underway. Experience with training data obtained from a full scope, high fidelity simulator at TVA’s Watts Bar Nuclear Power Plant, using simultaneously sampled values of the time records for some 22 to 27 variables for seven different accident transients (and a normal condition) shows that in all cases, the neural network was able to detect the transient before the plant tripped, even in the presence of 2% noise. This offers the opportunity to take mitigating actions (e.g., a power runback), where appropriate, either manually or automatically, to possibly prevent a plant trip. More recent work at Oak Ridge National Laboratory (Lin, Bartal and Uhrig, 1995) using data from the San Onofre-1 Nuclear Plant Simulator reported that this procedure showed high sensitivity for leaks in several plant systems (e.g., hot or cold leg loss of coolant accidents in the reactor coolant system, main steam line and feed water line breaks in or outside containment, etc.), using only reactor instrument readings input to an AI system. Benefits of this system. a) Almost immediate identification of plant transients (and associated faults), b) Ability to take mitigating action (e.g., power runback) before trip, if appropriate, and c) Identify and quantity very small leaks in the primary and secondary systems from monitoring control room instrumentation only. E. Increase in Thermal Power through Monitoring Reactor Core Parameters (DNBR Margin and Linear Power). The concept of increasing the licensed thermal output of the core through the use of a core monitoring system was approved in principle by the NRC when they licensed the Arkansas Nuclear One, Unit 2 plant (ANO-2) over two decades ago. (ANO-2, a two-loop Combustion Engineering Nuclear Power Plant, was licensed at a thermal power levels to produce 858 MWe compared to 830 Mwe for Maine Yankee and 839 MWe for St. Lucie Plants 1 and 2, all of which were virtually identical plants.) In

5

the ANO-2 case, a core protection calculator (CPC) utilized outputs from sensors located in the core to monitor the DNBR (departure from nucleate boiling regime) and thermal power per unit length in the fuel rod having the highest thermal output. Each additional Mwe output of the plant brings in additional revenues of about $150,000 per year with very little increase in operating and fuel costs. At 2.5c per KWe Hr, ANO-2 produced additional electricity that was worth over $6.1 million per year compared to Maine Yankee and $4.1 million per year per plant compared to each of the St. Lucie units. Recent work in South Korea using data from a KEPCO (Korean Electric Power Company) nuclear power plant indicates that it is possible to monitor thermal margin (i.e., the difference between the predicted DNBR and the limiting DNBR) using a neural network model (Kim, Lee, and Chang, 1992, 1993). The trained neural network model is utilized to predict the DNBR for a given set of operating conditions represented by the neural network inputs. The approach used is to train a neural network to map the plant variables being monitored to the DNBR as calculated by the computer code COBRA. The neural network is then trained over the range of input variables that was expected to occur during the fuel cycle. A statistical sensitivity analysis relating the DNBR to the various parameters indicated that the major parameters affecting DNBR of a PWR during plant operation were the core inlet temperature, the core power (or heat flux), the enthalpy rise peaking factor, the core inlet flow rate, and the system pressure. The DNBR obtained from the neural network using data not used in the training process showed that under steady state conditions, the results agreed with those obtained from COBRA calculations within +2.5% or better in almost all cases. The development of a similar but improved methodology (or the improvement of the Korean methodology), along with its extension to estimation of maximum linear thermal power along the fuel elements, would provide utilities with an opportunity to increase their thermal power without extensive plant modifications (subject, of course, to regulatory body approval). Benefits of this system: a) Ability to monitor thermal margin of the core, b) Justification for operating with a smaller thermal margin, thereby getting additional power form the same core, and c) Potential additional revenues of $150,000 per year for each additional KWe obtained from the plant. System Architecture. These five modules will be integrated into the conceptual design indicated in Figure 1. The system will receive data from a data acquisition system or the SPDS of a nuclear power plant, or from a nuclear power plant simulator. Data will be introduced simultaneously to all five modules and processed, to the extent possible, in real time. Plant status and other results will be displayed to the user through the graphical interface. III. HUMAN FACTORS, HUMAN MACHINE INTERFACE, AND SYSTEM TESTING Overview. INEEL's role in this project will focus on the human factors aspects of the instrumentation system, user requirements for the integration of artificial intelligence systems for nuclear plant monitoring and diagnostics, and testing of the prototype system. The planned coordination of INEEL and UT activities represents an effective integration of the capabilities of both organizations. The activities that are planned to address human factors, human-system interface, and testing requirements are briefly described in the following. Information Requirements Analysis. INEEL will systematically identify operating crew information requirements for nuclear plant monitoring and diagnostics. This information requirements analysis will ensure that the developed information and display system will effectively support enhanced crew performance. The analysis will be performed using the functional modeling approach that was developed by INEEL and applied to identify information requirements for severe accident management in a.

6

Command Module

Sensor Monitoring Transient Diagnosis Core Monitoring

Plant Data

Data Manager

User Interface

Efficiency Monitoring Inferential Sensing

Figure 1. General System Structure study for the Nuclear Regulatory Commission (NUREG/CR-5513). Subsequently, the method has been used at INEEL to assess information requirements for military aircraft, waste handling operations, and commercial air traffic control Interface Specification. Based on the results of the information requirements analysis, INEEL will develop the interface specification for the integrated information system. This effort will utilize the methods and capabilities developed at INEEL for the development of display systems for commercial power plant applications and the Advanced Test Reactor (ATR) at INEEL. This specification will provide detailed guidance to the UT personnel who will develop the interface. In addition, INEEL will support UT during the actual development of the interface in order to resolve human factors issues that arise. Test and Evaluation Plan. UT and INEEL will jointly develop a test and evaluation (T&E) plan. This plan will ensure that system testing is performed systematically, and that the conclusions reached from system testing are meaningful, accurate, and defensible. System Testing. UT and INEEL will jointly conduct the test and evaluation program and the analysis of the data collected. The final report will include an assessment of the effectiveness of the integrated information system for nuclear power plant monitoring and diagnostics, as well as identification of desired modifications for commercial application, both in the nuclear power field and in other domains such as fossil power plants, chemical plants, manufacturing plants, etc.

IV. STATUS OF TECHNOLOGIES INVOLVED A. Artificial Intelligence Technologies. Each of the artificial intelligence technologies to be applied to nuclear power plants in this project has been demonstrated to be technically feasible Some applications have been developed using MATLAB and SIMULINK, and the adaptation to this project will be straight forward. Others have been developed using a variety of commercial software (NeuralWare, ANSIM, AI-Ware, etc) and/or customized or specially written software, and the adaptation will be more time consuming. To the extent possible, systems will be modularized and isolated so that there is a minimum of potentially degrading interaction. Every methodology used will be reviewed as to how it can

7

be improved for the specific application in this project. Integration and modularization will be the major tasks because of the different nature of the various systems. The interface between the prototype system and the plant most likely will be through the Safety Parameter Display System (SPDS) or through the plant computer. B. Hardware. The latest personal computer technology will be used with 21- or 29-inch, high resolution monitors for the development work and for the prototype unit. It is anticipated that the prototype instrument will involve one very fast computer and one or possibly two 25" monitors for display. For this prototype, there will be no attempt to reduce the software system to hardware (microchips), because it is anticipated that the speed of the software system when converted to C or C++ will be more than adequate to meet the needs of the utility and process industries. However, for application in high performance systems (aircraft, missiles, etc.) the system could be implemented in customized microchips. V. APPLICATIONS OF SYSTEMS TO OTHER COMPLEX SYSTEMS UT is currently working on the application of autoassociative neural networks to instrument surveillance and calibration verification in the chemical process industry. Use of inferential measurements and neural network modeling and sensitivity analysis to iteratively improve plant efficiency has wide-scale application in the process and chemical industries as well as in fossil power plants. Applications of other features of this system such as inferential sensing and virtual instrumentation and instrument surveillance and calibration verification are also clearly appropriate. However, the application of transient analysis and diagnostics outside the nuclear power field would be dependent upon the availability of results of high fidelity simulator studies that have a degree of validity comparable to the results obtained from plant simulators used in the nuclear power industry. The Electric Power Research Institute has made available data from their Instrumentation and Control Center at the Kingston Fossil Plant are available to UT for the acquisition and interpretation of data. Although this laboratory is located at the site of a fossil power plant, it is de facto the research and development laboratory for instrumentation and control systems for nuclear power plants. TVA in cooperation with the EPRI I&C Center has instrumented its Kingston Unit #9 for thermodynamic measurements and has provided a modern distributed control and data acquisition system. The availability of data from this plant will be invaluable in adapting the prototype system to fossil power plants. It is expected that the technology developed for this project may have more general applicability across a broad range of industries. There is a general need to enhance the operational performance and safety of systems in the domains of commercial aviation, manufacturing, offshore oil, shipping, air traffic control, and chemical production. In many of these environments the effective integration of the human operators with automated systems is essential to optimize overall system performance. This project will generate experience to help assess the potential of performance gains that can be obtained through the effective integration of advanced monitoring and diagnostic systems.

VI. REFERENCES Y. Bartal, J. Lin and R. E. Uhrig "Nuclear Power Plant Transient Diagnostics Using Artificial Neural Networks that Allow 'Don't-Know' Classifications," Nuclear Technology, Vol. 110, No.3, pp 436-449, June 1995. E. Bartlett and R. E. Uhrig , “Nuclear Power Plant Status Diagnostics Using An Artificial Neural Network" Nuclear Technology, Vol.97, pp 272-281, March 1992.

8

Z. Guo, and R. E. Uhrig, "Using Modular Neural Networks to Monitor Accident Conditions in Nuclear Power Plants," Proceedings of the SPIE Technical Symposium on "Intelligent Information Systems, Application of Artificial Neural Networks III, Orlando, FL, April 20-24, 1992. Z, Guo and R. E. Uhrig, "Use of Artificial Neural Networks to Analyze Nuclear Power Plant Performance," Nuclear Technology, Vol.99, pp.36-42, July, 1992. J. W. Hines, D. J. Wrest, and R. E Uhrig, "Plant-Wide Sensor Calibration Monitoring," Proceedings of the 1996 IEEE International Symposium on Intelligent Control, Detroit, Ml, September 15-18, 1996. J. W. Hines, D.W. Miller and B. K. Hajek, "Merging Process Models with Neural Networks for Nuclear Power Plant Fault detection and Isolation," Proceedings of the 9th Power Plant Dynamics Control, and Testing Symposium, pp 54.01-12, Knoxville, TN, May 24-26, 1995. J. W. Hines, D. J. Wrest, and R. E. Uhrig,"Use of Autoassociative Neural Networks for Signal Validation," Proceedings of the NEURAP-97 Third International Conference on Neural Networks and their Applications, University of Aix-Marseille III, Marseilles, France, March 12-14, 1997. J. W. Hines, D. J. Wrest, and R. E. Uhrig, "Plant-Wide Sensor Calibration Monitoring," Proceedings of the 1996 IEEE International Symposium on Intelligent Control, Detroit, MI, September 15-18, 1996. A, Ikonomopoulos, L. H. Tsoukalas, and R. E. Uhrig, "Integration of Neural Networks with Fuzzy' Reasoning for Measuring Operational Parameters in a Nuclear Reactor," Nuclear Technology, Vol.104, October 1993.

K. Kavaklioglu and B. R. Upadhyaya, "Monitoring Feedwater Flow Rate and Component Thermal Performance of Pressurized Water Reactors by Means of Artificial Neural Networks," Nuclear Technology, Vol 107, July 1994. K.H. Kim, S. H. Lee, and S. H.. Chang, "Neural Network Model for On-Line Thermal Margin Estimation of a Nuclear Power Plant," Proc. of the Second Internarional Forum, Expert Systems , 1992. K. H. Kin, S. H. Chang, and S. H. Lee, “Pressurized Water Reactor core Parameter Predictions using an Artificial Neural Network,” Nuclear Science and Engineering, Vol. 115, pp 152-163, 1993. R.E. Uhrig, "Use of Neural Networks in Nuclear Power Plant Diagnostics," Proceedings of the In. Conf. on Availability Improvements in Nuclear Power Plants, Madrid, Spain, April 10-14, 1989. R. E. Uhrig and L. H. Tsoukalas, “Soft Computing Technologies in Nuclear Engineering Applications,” Advances in Nuclear Technology, (in press), 1998.

9