The work deals with the use of artificial neural networks for the monitoring and diagnosis of components and systems in nuclear power plants. The study has been conducted in two phases. First, the mechanical behavior of rotting machinery has been studied and a neural network system has been developed for the detection of faults in rolling element bearings. A monitoring technique for trending has also been proposed, which models the relationship among sensors in the machinery, and permits easy identification of changes in operating states.
The second phase of the study is related to the problem of the vibration of the internals of PWR reactors. The primary concerns are the detection of contacts between the core barrel and the reactor vessel, and the monitoring of the presence of these contacts since they may provoke fatigue in the structure. To this aim, a neural-bases system was designed which is able to discriminate among the different modes of vibration of the reactor internals using features extracted from the neutron noise power spectral densities. One Graduate Research Associate spent seven months at the EdF Research Center in Chatou (near Paris) working with the SINBAD data base which contains vibration data for the core of 28 almost identical 900 MWe PWR nuclear power plants.
Title: Diagnostic Problems Using Fuzzy Variables
The objective of this project is to develop a method of diagnosing problems in complex systems by relating the symptoms with the various faults. Since a given symptom maybe associated with several different faults, there is matrix type relationship between symptoms and faults. Furthermore, the degree of the symptoms and the degree of the faults are not directly related. This project utilizes fuzzy set theory to relate the symptoms and the faults. Furthermore, it deals with the individual symptoms and faults as fuzzy variables. Data from NUREG-1150, the Nuclear Regulatory Commission's probabilistic risk assessment for several nuclear power plants is being used to demonstrate the advantages of this approach.
Title: Application of Neural Networks to Monitoring Check Valves for Operability
Typical power plants have 200 to 300 check valves of which 15% are important to safety. Since these valves are self-actuating and are sealed, with no external indication of their operability, the NRC has required that some of these valves be opened and inspected to assure that they will operate when needed. Not only is such an inspection process expensive in terms of finances and radiation exposure, it is often counter productive in that working valves are sometimes damaged during the inspection. The goal of this project is to develop a means of inspecting these valves for operability using acoustical signals taken from two accelerometers places dome distance apart on the outside body of the valve.
The process involves using the spectra of two signals to model the interrelationship between these two accelerometer signals by training a neural network when the valve is known to be operating properly. Then the neural network is used in a monitoring mode with one signal as the input to the neural network. A comparison is made between the other signal and the corresponding signal predicted by the neural network. It has been found that mechanical problems inside the valve disturb the relationship between the two external signals. Therefore, a significant difference between the signal predicted by the neural network and actual signal indicates that the valve should be opened, inspected and repaired, if necessary.
Title: Use of Neural Networks Trained on Nuclear Power Plant Simulator Data to Diagnose faults in Nuclear Power Plants
This program in now in its third generation. Two Ph.D. Dissertations have been completed in developing this methodology. The methodology involves using data gathered from nuclear power plant training simulators (typically 20 to 30 variables sampled 2 to 5 times per second) during transients deliberately introduced into the plant as the input to the neural network. This approach has been shown to work satisfactorily and is generally able to diagnose the transient well before the plant trips off-line, thereby giving the operator the opportunity to take mitigating action, if appropriate.
The goal of the current work is threefold. First, the diagnostic system is being made more robust by introducing variations under different operation conditions (power level, boron concentration in the coolant, heat sink temperature, different equipment out of service, degraded heat transfer components due to fouling, etc.). The second objective is to find a way to identify the onset of a transient at the earliest possible time and to separate this indication from the normal fluctuations that occur in a nuclear power plant during normal day-to-day operation. Third, a method of quantitatively evaluating the magnitude of leaks from nuclear systems over the range from 0.1 to 1000 gallons per minute is being investigated.
The earlier studies utilized data from TVA's Watts Bar simulator for seven transients. Now we have access to the San Onofre Plant #1 simulator (recently purchased by Ryan Nuclear Co., when San Onofre Plant #1 shut down). This will allow us to run literally hundreds of variations of the 30 to 35 transients normally associated with the operation of nuclear power plants.
Title: Inferential Sensing
An inferential sensing system is an instrumentation system which infers values of complex process variable by integrating information from multiple sensors. They often incorporate neural network process for process modeling and can provide estimates of process variables that are usually measured off-line or through analytical laboratory instruments such as composition of fluids in a mixture. Such a system is especially useful in processes where direct instrumentation for monitoring the process is not possible, thereby precluding closed-loop control. Attractive features of inferential sensors embodies a process model on which sensitivity analysis can be executed. The result of the sensitivity analysis can be used to derive the control laws and the means of optimizing the process. Sensor failure detection is extremely useful in the process control where an unchecked failure can be expensive and potentially dangerous. The obvious application to nuclear power plants is for indirect measurement of reactivity which must be calculated from plant and core properties. The reactivity computer at Chernobyl was running minutes behind real-time at the accident. A neural network that has been trained using the results of calculations that cover the operating range of the input variables can provide instantaneous evaluation of reactivity. There are many other quantities in a nuclear power plant (e.g., DNBR--deviation from nucleate boiling regime, flow valve positions, etc.) That can be evaluated using inferential measurements.
Title: On-line Thermodynamics Modeling and Optimization
The major thrust of this program is to develop methods by which the inherent modeling capability of neural networks can be adapted to on-line, real time monitoring with iterative optimization to provide guidance to the operators (or to automatic control systems if safety systems are not involved) on improving heat rate. Increasing the robustness of the neural network model to deal with a wide spectrum of plant conditions that might be involved over an annual cycle of temperatures or the changes in neutronic conditions over a fuel cycle is objective. The possibility of using individual "modular" neural networks for each control system with only the most important input variables for each control system determined by an optimization technique (sensitivity analysis or genetic algorithms) will be investigated. An expert system, probably using fuzzy variables, will be coupled to the system of neural networks, to coordinate the outputs of the modular neural networks. Generally, it is expected that the data available at the SPDS (safety parameter display system) and/or the plant computer could provide data needed for modeling the thermodynamics of the plant. Testing on a full-fidelity training simulator will probably be necessary before proceeding with testing on a nuclear power plant. Since such a heat rate monitor might be equally useful for fossil power plants, this provides an alternative for testing.
Another issue that needs to be addressed is whether the improvements made possible through such a system warrant the installation of a sophisticated system with neural networks, clustering algorithms, sensitivity coefficients, expert systems, etc. A preliminary cost effectiveness evaluation will be made to determine the value of this heat rate monitor, Since an increase in efficiency of 0.1% in a 1200 MWe power is worth about $500,000 per year at 5 cents per KWe-Hr, it is anticipated that even small improvements in efficiency would be very cost effective.
Title: Monitoring and Verification of Nuclear Power Plant Instrumentation
The work being carried out here is an integration and application of neural network methodologies already being investigated at the University of Tennessee. The objective is to determine the feasibility of using one or more neural network methodologies to monitor the calibration and verify the proper operation of commercial nuclear power plant process instrumentation. It is the intent of this program that the method should be capable of identifying the deviation of a single sensor (e.g., due to drift or failure) through its interrelation with the outputs of other sensors. In the process of accomplishing these goals, it is expected that the project will provide an understanding of the behavior of neural networks and related artificial intelligence (soft computing) methodologies when measuring and monitoring the many variables of a complex system so that these methodologies are used correctly and effectively. With application-specific microchips available today, the proposed systems could be imbedded within a sensor or the relevant data processing system which could provide the ability to operate on-line in "real time."
Two different neural network technologies are being investigated as a means of accomplishing the above goals; autoassociative neural networks and an array of modular inferential neural networks. In the autoassociative neural network monitoring system, the neural network behaves as an associative memory into which a large number of complex relations are stored by the training process. The ability to obtain the original pattern relationships from the output when the input pattern relationships are distorted (i.e., one of the input values has changed) gives, in principle, the autoassociative neural network its unique capability to serve as a system-wide monitor. Sequential probability ratio tests (SPRT) comparing actual measurements with projected values from the neural networks will be used to identify deviations from normal operation.
In each modular inferential neural network, a single output variable is estimated through its interrelationship with several other measured plant variables. The choice of variables to use as inputs to the inferential neural network will be determined by an appropriate methodology, such as principle component analysis, genetic algorithms, sensitivity analysis, etc.). With an array of perhaps 15 modular networks, each trained to predict a single variable over an appropriate range of the input variables, the comparison of the predicted with the actual output variables will provide a close monitoring of all the variables involved. It is inevitable that plant variables will be used as inputs to several individual modular neural networks. Therefore, if that single measured variable drifts, it will affect the outputs of several individual neural networks. A change in the behavior of the outputs of the neural networks having that input compared with the behavior of the neural networks not having that variable as an input can confirm abnormal behavior of that sensor or channel.