Signal Validation Abstracts from Publications (Co)Authored by Dr. Holbert
Other Keywords: Fault Detection and Isolation, Sensor Validation.

Fuzzy Associative Memories for Instrument Fault Detection

A. Sharif Heger, Keith E. Holbert, A. Muneer Ishaque

Abstract

A fuzzy logic instrument fault detection scheme is developed for systems having two or three redundant sensors. In the fuzzy logic approach the deviation between each signal pairing is computed and classified into three fuzzy sets. A rule base is created allowing the human perception of the situation to be represented mathematically. Fuzzy associative memories are then applied. Finally, a defuzzification scheme is used to find the centroid location, and hence the signal status. Real-time analyses are carried out to evaluate the instantaneous signal status as well as the long-term results for the sensor set. Instantaneous signal validation results are used to compute a best estimate for the measured state variable. The long-term sensor validation method uses a frequency fuzzy variable to determine the signal condition over a specific period. To corroborate the methodology synthetic data representing various anomalies are analyzed with both the fuzzy logic technique and the parity space approach.

Annals of Nuclear Energy, Vol. 23, No. 9, pp. 739-756, 1996.


A Hybrid Learning System for Sensor Fault Detection and Isolation

A. Sharif Heger, Keith Holbert

Abstract

Optimal control of a complex dynamic system such as a power plant is dependent on the validity of sensors providing information about the process state. To support better diagnostics and improve plant performance, a hybrid fuzzy-logic and crisp system for sensor fault detection and isolation (FDI) is under investigation. This research uses fuzzy logic to transform linguistically expressed signal analysis principles into a classification rule-base for signal failure detection and identification. The first part of this paper recounts the implementation of a fuzzy-logic based method for performing signal validation on redundant sensors; the second part describes current research at extending this work to develop a hybrid learning system for FDI.

Soft Computing with Industrial Applications, Vol. 5, pp. 295-300, Proceedings of the World Automation Congress, Montpellier, France, May 28-30, 1996.


Fuzzy Logic for Power Plant Signal Validation

Keith E. Holbert, A. Sharif Heger, A. Muneer Ishaque

Abstract

A fuzzy logic signal validation scheme is developed for systems having two or three redundant sensors. In the fuzzy logic approach the deviation between each signal pairing is computed and classified into three fuzzy sets. A rule base is created allowing the human perception of the situation to be represented mathematically. Fuzzy associative memories are then applied. Finally, a defuzzification scheme is used to find the centroid location, and hence the signal status. Real-time signal analyses are carried out to evaluate the instantaneous signal status as well as the long-term results for the sensor set. Instantaneous signal validation results are used to compute a best estimate for the variable being measured. The long-term signal validation method uses a frequency fuzzy variable to determine the signal status over a specific period. To validate the methodology synthetic data representing various anomalies are analyzed with both the fuzzy logic technique and the parity space approach.

Proceedings of the Ninth Power Plant Dynamics, Control & Testing Symposium, pp. 20.01-20.15, Knoxville, TN, May 1995.


A Neural Network Realization of Linear Least-Square Estimate for Sensor Validation

Xianzhong Wang, Keith E. Holbert

Abstract

In applications of artificial neural networks for sensor validation and fault diagnosis, a plant model or process knowledge can help overcome the incompleteness of training data and improve the generalization of neural networks. In this research, a steady state plant model with Gaussian noise distribution is directly incorporated into the neural network. A three-layer artificial neural network is constructed to perform the sensor validation. The first hidden layer of the network is made to output the least mean-square estimate of the state variable vector, i.e., the first layer is a neural network realization of a least mean-square estimator. It is achieved by making the nodes linear and assigning the proper weight to each connection with the input units (measurements). The activation functions of the nonlinear output nodes are determined according to the Gaussian noise distribution and the plant model. The output nodes give a probabilistic description of the residuals' deviation from zero for the corresponding measurements.

Proceedings of the Ninth Power Plant Dynamics, Control & Testing Symposium, pp. 15.01-15.15, Knoxville, TN, May 1995.


Redundant Sensor Validation by Using Fuzzy Logic

Keith E. Holbert, A. Sharif Heger, Nahrul K. Alang-Rashid

Abstract

This research is motivated by the need to relax the strict boundary of numeric-based signal validation. To this end, the use of fuzzy logic for redundant sensor validation is introduced. Since signal validation employs both numbers and qualitative statements, fuzzy logic provides a pathway for transforming human abstractions into the numerical domain and thus coupling both sources of information. With this transformation, linguistically expressed analysis principles can be coded into a classification rule-base for signal failure detection and identification.

Nuclear Science and Engineering, Vol. 118, No. 1, pp. 54-64, 1994.


Empirical Process Modeling Technique for Signal Validation

K. E. Holbert, B. R. Upadhyaya

Abstract

Some techniques for fault detection involve the comparison of measured process signals with independent estimates. The prediction of process variables can be achieved either by physical or empirical modeling of a plant subsystem. An automated procedure for generating empirical process models is developed here. Independent prediction of critical signals is required for consistency checking of instrument outputs, for their degradation monitoring and for isolating common-mode failures. The steady-state empirical models are developed using data from different steady-state conditions. Signal anomaly is identified by comparing the error between the model-based prediction and the actual measurement with a fuzzy function (curve) utilizing the signal tolerance as a threshold. In the event a signal is declared as failed, the predicted estimate is used as input to a control/safety system or for display to an operator. Application of the methodology to signal validation using operational data from a commercial PWR and the EBR-II is presented.

Annals of Nuclear Energy, Vol. 21, No. 7, pp. 387-403, 1994.


Sensor Validation Using Fuzzy Logic (Fault Detection in the Instrumentation System)

Keith E. Holbert

Abstract

Signal validation is used to identify faulty process measurements. The incentives for performing sensor validation reside in both safety concerns and the economic returns possible. A limitation of current crisp fault detection techniques is that data from at least two domains influence the subsequent classification actions. The first domain, numeric in nature, originates from the process instrumentation. The other source stems from the human operators who interact with the complex system. This source is, without exception, expressed in linguistic terms; it is imprecise and subjective. It is imbedded in the operator's mind and represents his epistemic state. This state is formed by the operator's experience and his conceptualization of the complex system. For example, in approaching the task of redundant sensor validation from an operator's perspective, expressions such as "the sensors agree most of the time" and "that sensor is consistently biased low from the others" are encountered. From such an analysis a sensor is judged as "valid," "failed," or "suspect."

The reason for using fuzzy logic as a tool for sensor validation is its ability to transform information from the linguistic domain to a mathematical domain coupling with sensor data and then transformation of its result back into the linguistic domain for presentation. The inputs to the fuzzy logic evaluator consist of the signal measurements themselves. The fuzzy logic approach breaks the hard thresholds used in crisp techniques into less rigidly fixed regions of "small," "medium," and "large" deviations. The output from the technique is a statement of the validity of each input signal. The use of fuzzy logic transitions the boundary between valid and failed with a suspect regime that may aid in the prevention of catastrophic failure (i.e., preventive maintenance).

Fuzzy logic is used to transform linguistically expressed signal analysis principles into a classification rule-base for signal failure detection and identification. The rule-base of the fuzzy logic approach is generic in the sense that the rules are oblivious to the particular state variable being measured. The compactness of the rule-base and the simplicity with which the deviations and frequency fuzzy variables are applied with the rules allows signal processing in real-time. The use of fuzzy logic for sensor validation is explored, and examples are presented.

Electric Power Research for the 90's, Proceedings of the Fourth Annual Industrial Partnership Program Conference, April 28, 1994.


Neural Networks for Signal Validation in Nuclear Power Plants

Keith E. Holbert

Abstract

Signal validation is an important part of process control and monitoring. To date many techniques have been developed, and are being developed, for the purpose of verifying signal integrity. The objective of this investigation is to determine the feasibility of using a feed-forward backpropagation neural network in a signal fault detection capacity.

Software is developed to simulate neural networks having both one and two hidden layers. In addition, software is developed for transforming the signal data into a workable format. The application of the network learning algorithm to operational power plant data necessitates the development of an automated method to generate "bad" data points with which the network can make contrasts against "good" data points during the learning stage. Data preprocessing includes normalizing the input from zero to unity using the transducer range as a reference. The network has an input and output node corresponding to each of the process signals. Ideally, the output is unity or zero indicating the condition, valid or faulty respectively, of the corresponding input signal.

Sensor data are taken from various subsystems within an operating nuclear power plant for use in this investigation. Both single and double layer neural networks are trained to validate each of the data sets. During network testing the output errors are classified as either missed or false alarms in order to evaluate the network performance. The results of this investigation are given in the presentation.

Electric Power Research for the 90's, Proceedings of the Second Annual Industrial Partnership Program Conference, March 9, 1992.


Process Hypercube Comparison for Signal Validation

Keith E. Holbert

Abstract

The optimal control and safe operation of a nuclear power plant requires reliable information concerning the state of the process. Signal validation is the detection, isolation, and characterization of faulty signals. Properly validated process signals are beneficial from the standpoint of increased plant availability and reliability of operator actions.

A signal validation technique utilizing a process hypercube comparison (PHC) was originated during this research. The hypercube is merely a multidimensional joint histogram of the process conditions. The hypercube is created off-line during a learning phase using operational plant data. In the event that a newly observed plant state does not match with those in the learned hypercube, the PHC algorithm performs signal validation by progressively hypothesizing that one or more signals is in error. This assumption is then either substantiated or denied. In the case where many signals are found to be in error, a conclusion that the process conditions are abnormal is reached. The global data base contained within the hypercube provides a best estimate of the process conditions in the event a signal is deemed failed.

The hypercube signal validation methodology was tested using operational data from a commercial pressurized water reactor (PWR) and the Experimental Breeder Reactor II (EBR-II). This research was part of a larger project aimed at the development of a comprehensive signal validation software system for application to nuclear power plants.

IEEE Transactions on Nuclear Science, Vol. 38, No. 2, pp. 803-811, 1991.


An Integrated Signal Validation for Nuclear Power Plants

Keith E. Holbert, Belle R. Upadhyaya

Abstract

The optimal control and safe operation of a nuclear power plant requires reliable information concerning the state of the process. Signal validation is the detection, isolation, and characterization of faulty signals. Properly validated process signals can provide increased plant availability and reliability of operator actions.

A comprehensive signal validation software system has been developed for application to nuclear power plants. This system combines some previously established fault detection methodologies as well as some newly developed modules. The techniques have been implemented in a modular architecture that allows for the addition or removal of signal validation "modules" as deemed necessary. Intramodule confidence factors describing the validity of a given signal are derived using fuzzy membership functions. A final evaluation of signal status is made by the system executive based on results from each signal validation module. To make reliable decisions in this parallel system, a positive decision maker was developed. The hypercube signal validation methodology and the comprehensive system were tested using operational data from both a commercial pressurized water reactor and the Experimental Breeder Reactor II.

Nuclear Technology, Vol. 92, No. 3, pp. 411-427, 1990.


Last updated: January 29, 1997
Return to Dr. Holbert's Publications