EPRI Automated Analysis of Bobbin CoilProbe Eddy Current Data

Automated Analysis of Bobbin Coil Probe Eddy Current Data 1002785 Final Report, December 2002 EPRI Project Manager J. Benson EPRI • 3412 Hillview Avenue, Palo Alto, California 94304 • PO Box 10412, Palo Alto, California 94303 • USA 800.313.3774 • 650.855.2121 • [email protected] • www.epri.com http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES THIS DOCUMENT WAS PREPARED BY THE ORGANIZATION(S) NAMED BELOW AS AN ACCOUNT OF WORK SPONSORED OR COSPONSORED BY THE ELECTRIC POWER RESEARCH INSTITUTE, INC. (EPRI). NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, THE ORGANIZATION(S) BELOW, NOR ANY PERSON ACTING ON BEHALF OF ANY OF THEM: (A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I) WITH RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM DISCLOSED IN THIS DOCUMENT, INCLUDING MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON OR INTERFERE WITH PRIVATELY OWNED RIGHTS, INCLUDING ANY PARTY'S INTELLECTUAL PROPERTY, OR (III) THAT THIS DOCUMENT IS SUITABLE TO ANY PARTICULAR USER'S CIRCUMSTANCE; OR (B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER (INCLUDING ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVE HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOUR SELECTION OR USE OF THIS DOCUMENT OR ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM DISCLOSED IN THIS DOCUMENT. ORGANIZATION(S) THAT PREPARED THIS DOCUMENT Michigan State University ORDERING INFORMATION Requests for copies of this report should be directed to EPRI Orders and Conferences, 1355 Willow Way, Suite 278, Concord, CA 94520, (800) 313-3774, press 2 or internally x5379, (925) 609-9169, (925) 609-1310 (fax). Electric Power Research Institute and EPRI are registered service marks of the Electric Power Research Institute, Inc. EPRI. ELECTRIFY THE WORLD is a service mark of the Electric Power Research Institute, Inc. Copyright © 2002 Electric Power Research Institute, Inc. All rights reserved. http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 CITATIONS This report was prepared by Michigan State University Engineering Building, Room 2120 East Lansing, MI 48824 City, State Zip Principal Investigator S. Udpa This report describes research sponsored by EPRI. The report is a corporate document that should be cited in the literature in the following manner: Automated Analysis of Bobbin Coil Probe Eddy Current Data, EPRI, Palo Alto, CA: 2002. 1002785. iii http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 REPORT SUMMARY This report provides a summary of ongoing research to develop algorithms for performing automated analysis of eddy current test data. The research focus is on analysis of bobbin coil data for detecting degradation in steam generator tubes. Background Automated eddy current data analysis systems have been available for more than a decade to provide rapid detection of degradation in steam generator tubing. EPRI published an assessment Assessment of Automated Eddy Current Data of commercial data analysis software in 2002: Analysis Technology for Steam Generator Tubing, Bobbin Coil Probe Data (EPRI report 1003140). That report stated none of the three systems evaluated were successful in detecting degradation in all 21 damage mechanism categories and that overcall rates for the once through steam generator (OTSG) data sets were much higher than desired. Limitations in detecting degradation or inefficiencies resulting from large numbers of overcalls have limited use of automated data analysis systems at some plants. Objectives To develop algorithms that will automatically analyze bobbin coil eddy current data and identify degradation in steam generator tubes; to achieve values for probability of detection (POD) greater than 90% for all known degradation categories; and, to achieve degradation overcall rates lower than achieved by available commercial systems. Approach The project team was provided with steam generator bobbin coil eddy current inspection data to develop algorithms for automatically detecting and classifying degradation in steam generator tubes. Along with this set of “training” data, expert opinion analysis results from two independent qualified data analysts also were provided. This same set of training data had been previously used to assess the capabilities of commercially available automated data analysis systems. Initial algorithm development will focus on OTSG data, which has traditionally been the most challenging for automated data analysis. Inclusion of data from Westinghouse and Combustion Engineering (CE) steam generators also is planned. Once the algorithms produce superior results on the training data compared to results from commercial systems, a “test” data set will be processed by the newly developed algorithms, without the help of expert opinion results. Following a successful performance demonstration, the newly developed algorithms can either be incorporated into a stand-alone system or be added as a separate data analysis option to existing commercial data analysis systems. v Results http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Results Previous EPRI work included an assessment of three commercial automated data analysis systems on a common set of field bobbin coil data. The assessment included training data from 67 OTSG tubes. These same 67 tubes were included in the present EPRI algorithm development project aimed at improving the capabilities of automated data analysis. Development of a multistage classification algorithm has produced promising automated data analysis results. The automated data analysis system consists of two stages: signal preprocessing (to remove noise and low-frequency trends) and signal classification (uses rule-based algorithms and statistical models that account for inherent variability in real-world systems). The EPRI-developed algorithm detected 96% of the defects in the OTSG training data, which is comparable to results achieved by the three commercial systems (range of detection: 93% to 98%). The EPRI-developed algorithm defect overcall rate of 9 per tube represented a significant improvement over the 13 to 34 overcalls per tube by the three commercial systems. EPRI Perspective In 2000, EPRI initiated a project to develop software algorithms to perform automatic analysis of bobbin coil eddy current data. To date, the project has resulted in extremely promising results, and it is expected that the algorithms developed will provide improved capabilities (higher POD and lower false call rates) than are currently available from commercial systems. Following a successful demonstration of the algorithms, they could be used either in a standalone system or as an analysis option from within another data analysis software product. Whether used alone or in parallel with another independent automatic analysis software product, utilities would benefit from EPRI-developed automatic analysis software. Benefits would include cost savings associated with a reduction in data analyst staffing levels, more consistent data analysis leading to increased POD, and a shorter inspection duration. Currently, the newly developed automated data analysis software is being prepared to process field data from over 600 OTSG tubes. Results of the software validation on this test data set will be used to identify where algorithm improvements may be needed. Similar algorithm validation is planned in 2003 for field data from hundreds of tubes from Westinghouse- and CE-designed steam generators. Keywords Steam generators Automatic data analysis Eddy current Bobbin probe vi http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 ABSTRACT Automated eddy current data analysis systems for detection of steam generator tube degradation can provide significant benefits to utilities. Potential benefits may include: 1. Cost savings by reducing manpower and equipment requirement needs, 2. Schedule savings by: – – reducing the impact of data analyst shortages during peak outage periods providing analysis results at rates equal to or less than the rates for data acquisition 3. Reliability improvements by providing consistent, repeatable and accurate inspection results This report provides a summary of the status of ongoing research to develop algorithms for performing automated analysis of eddy current test data. The focus of the research is on analysis of bobbin coil data for detection of degradation in steam generator tubes. Chapter 1 of the report begins with a discussion of the design of steam generators in nuclear power plants and a summary of degradation mechanisms that have occurred in steam generator tubing. The principles of eddy current testing are briefly discussed. The initial step of the data analysis process, signal preprocessing, is described in Chapter 2. A series of preprocessing routines are used to process the raw data and identify those data segments that require further analysis. Various types of data segmentation, filtering, de-noising and thresholding are part of the signal preprocessing step. Chapter 3 addresses the multistage approach of data processing and signal classification aimed at reducing the number of overcalls while maintaining a high degradation detection rate. A combination of rule bases and statistical classifiers are used to eliminate the non-defect indications systematically while retaining the defect and dent indications at each stage. In addition, certain classes of defects are handled separately. Volumetric indications, like wear and impingement, merit their own processing routines. Separate algorithms for wear and impingement detection were also developed. The results of various stages of the automated data analysis process are provided throughout the report, and summarized in Chapter 4 following the completion of all data analysis stages. The data used to demonstrate the effectiveness of the various algorithms consists of a set of field data from once through steam generators. The results of the algorithm are compared to the “expert opinion” results that were verified by two independent Qualified Data Analysts. vii http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 CONTENTS 1 INTRODUCTION ................................................................................................................. 1-1 Steam Generator in Nuclear Power Plant........................................................................... 1-1 Principles of Eddy Current Testing and its Application in Steam Generator Tube Inspection........................................................................................................................... 1-2 Multifrequency Eddy Current Techniques ...................................................................... 1-4 Bobbin Coil Probe.......................................................................................................... 1-5 Research Objectives .......................................................................................................... 1-6 2 SIGNAL PREPROCESSING ............................................................................................... 2-1 Data Segmentation............................................................................................................. 2-2 Adaptive Filtering................................................................................................................ 2-3 Adaptive Filtering using the NLMS Algorithm ................................................................. 2-3 Wavelet Shrinkage De-Noising ...................................................................................... 2-5 Zero-phase High Pass Filter............................................................................................... 2-5 Dynamic Thresholding (Neyman-Pearson Detector)........................................................... 2-7 Neyman-Pearson Detector ............................................................................................ 2-7 Moving Average Filter........................................................................................................2-12 Distance Threshold ...........................................................................................................2-13 Results of the Preprocessing Module ................................................................................2-14 3 MULTISTAGE SIGNAL PROCESSING AND CLASSIFICATION........................................ 3-1 Magnitude Thresholds........................................................................................................ 3-1 Calibration Curve based Phase Thresholds........................................................................ 3-1 Rule Base I......................................................................................................................... 3-3 Rule Base II........................................................................................................................ 3-8 Hidden Markov Models......................................................................................................3-10 Eddy Current Classification...........................................................................................3-11 Impingement Classifier ......................................................................................................3-12 ix Wear Identification.............................................................................................................3-14 4 SUMMARY AND CONCLUSIONS....................................................................................... 4-1 5 REFERENCES .................................................................................................................... 5-1 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 x LIST OF FIGURES http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-1 Steam Generator in Nuclear Power Plant.............................................................. 1-1 Figure 1-2 Eddy Current Generation and Flow in a Conducting Specimen.............................. 1-3 Figure 1-3 Impedance-plane Trajectory of a Coil over a Conducting Non-ferromagnetic Test Specimen [3] ........................................................................................................... 1-4 Figure 1-4 Differential Bobbin Probe, Comprising two Coaxial Air Core Probes in a Tube ...... 1-6 Figure 1-5 Schematic of a differential bobbin probe scanning a heat exchanger tube ............. 1-7 Figure 1-6 Typical differential probe signals. The lower traces are strip chart displays of the respective imaginary (left) and real (right) components of the eddy current probe signals............................................................................................................................. 1-8 Figure 1-7 Multifrequency Bobbin Coil Probe Eddy Current Signals........................................ 1-9 Figure 1-8 Eddy Current Data Analysis System .....................................................................1-10 Figure 2-1 Overall Structure of the Preprocessing Algorithm................................................... 2-1 Figure 2-2 Entry and Exit point signals in the tube data........................................................... 2-2 Figure 2-3 Block diagram of the adaptive filtering algorithm. ................................................... 2-3 Figure 2-4 Schematic of the NLMS adaptive noise rejection system used for minimizing noise in eddy current data. .............................................................................................. 2-4 Figure 2-5 (a) Original raw data, (b) High pass filtered (DCT) data.......................................... 2-6 Figure 2-6 Improved Dynamic Thresholding Algorithm........................................................... 2-9 Figure 2-7 (a) Original raw data (Vertical, mix channel), (b) Output of Dynamic Thresholding (possible defect locations).........................................................................2-10 Figure 2-8 (a) Illustration of the output potential defect points of the Dynamic Thresholding algorithm, (b) Zoomed in view of the defect region....................................2-11 Figure 2-9 (a) Input to Moving Average filter, (b) Output of Moving Average Filter .................2-12 Figure 2-10 (a) Locating the minima in the vertical component, (b) Corresponding minima represented in the impedance plane ..................................................................2-13 Figure 3-1 Block diagram of the Multistage classification module............................................ 3-2 Figure 3-2 The ideal and practically used phase calibration curves......................................... 3-2 Figure 3-3 Phase calibration curves for all four channels of one of the calibration tubes R999C999G003 .............................................................................................................. 3-3 Figure 3-4 The IPT of a DENT indication in all four channels .................................................. 3-4 Figure 3-5 The IPT of an OD defect in all four channels.......................................................... 3-5 Figure 3-6 The IPT of an ID indication in all four channels ...................................................... 3-6 Figure 3-7 Overall approach of Rule Base I. ........................................................................... 3-7 xi Figure 3-8 Scatter plot of the variance of the 400 kHz channel for one plant........................... 3-9 Figure 3-9 Cross correlation between the horizontal components vs. the cross correlation between the vertical components (200 kHz & 400 kHz). ...............................3-10 Figure 3-10 Flowchart for eddy current signal classification using HMMs (a) Training and (b) Testing ......................................................................................................................3-11 Figure 3-11 Vertical component and IPT plot of an impingement ...........................................3-13 Figure 3-12 MIX channel vertical and horizontal components of two support plate signals ith t 3 15 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 without wear. ..................................................................................................................3-15 Figure 3-13 MIX channel vertical and horizontal components of two support plate signals with wear. .......................................................................................................................3-16 xii LIST OF TABLES Table 2-1 Data distribution in the EPRI OTSG training database 2-14 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Table 2 1 Data distribution in the EPRI OTSG training database. ..........................................2 14 Table 2-2 Results of the Preprocessing Module.....................................................................2-14 Table 2-3 Analysis of missed indications in the EPRI OTSG training database. .....................2-15 Table 3-1 Summary of results after magnitude & phase thresholding, followed by Rule Base I.............................................................................................................................. 3-7 Table 3-2 Summary of missed flaws/dents in Rule Base I....................................................... 3-8 Table 3-3 Results of Applying Rule Base II. ...........................................................................3-10 Table 3-4 Performance of the HMM on the EPRI OTSG training database. ...........................3-12 Table 3-5 Results of the IMPINGEMENT CLASSIFIER on all four plants...............................3-14 Table 3-6 Summary of the wear classifier. .............................................................................3-17 Table 4-1 Overall summary of the OTSG classification algorithm............................................ 4-2 Table 4-2 Summary of missed flaws. ...................................................................................... 4-2 xiii http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 1 INTRODUCTION Steam Generator in Nuclear Power Plant http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 H h b di i fi d i i l di i h i l Heat exchanger tubes are used in a variety of industries, including, power stations, petrochemical plants, oil refineries, air conditioning and refrigeration units for transferring heat to the fluid circulating outside the tube. A steam generator or heat exchanger unit used in nuclear power plants, shown in Figure 1-1, transfers heat from the primary loop to the secondary to produce steam, which is used to run the turbines. Typically these tubes are made of Inconel and, for one steam generator design, are approximately 7.5m high with an internal diameter of 15.5mm and 1mm wall thickness. The tube bundle is supported by ferromagnetic support structures, which are distributed along the length of the tubes. Figure 1-1 Steam Generator in Nuclear Power Plant The steam generator tubes are continuously exposed to harsh environmental conditions including high temperatures, pressures, fluid flow rates and material interactions resulting in various types of degradation mechanisms such as mechanical wear between tube and tube supports, outer 1-1 Introduction diameter stress corrosion cracking (ODSCC), pitting, thinning, primary water stress corrosion cracking (PWSCC), and inter granular attack (IGA). Tube degradation can progress completely through the tube wall, thereby contaminating the fluids on the secondary side of the steam generator. It is critical that the primary coolant, which is radioactive, does not leak into the secondary side. Consequently the steam generator tubes in nuclear power plants need to be inspected periodically for degradation. Historically, steam generator tube inspection has been a challenging issue. There have been numerous cases of unscheduled plant shutdowns in the past, which typically cost $500,000 a day. Hence there is a strong economic incentive to develop reliable nondestructive evaluation (NDE) methods for steam generator tube inspection. Visual examination and ultrasonic techniques have limited use as they are very slow and time consuming. These obstacles have led the way to a widespread use of eddy current techniques for the inspection of non-ferrous tubing, particularly http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 p y q p g p y in the nuclear power industry. Eddy current inspection has proven to be both fast and effective in detecting and sizing most of the degradation mechanisms that occurred early in the life of first generation steam generators. However, as the nation’s steam generators have aged, newer and much more subtle forms of degradation have appeared that require more intelligent application of eddy current tests. Conventionally, eddy current data analysis is carried out by human analysts. Normally, during inspection, signals from multiple channels, frequencies and probe types are recorded. Data analysts use their experience as they review the shape of Lissajous patterns and phase of the signal in each channel to make an assessment of the tube condition. Through the use of multifrequency eddy current systems, modern equipment is now capable of acquiring the necessary data to detect and correctly diagnose indications of tube degradation. Consistent and reliable analysis techniques are required to achieve improved detection, classification and characterization results. Human analysis, apart from being slow, is often inconsistent. Inspection results are often not consistent with prior inspection and/or with other analysts. Thus, there is a need for automated signal classification systems that can provide accurate and consistent signal interpretation. Principles of Eddy Current Testing and its Application in Steam Generator Tube Inspection The basic principle underlying eddy current testing can be illustrated with a simple arrangement shown in Figure 1-2 [1]. When a coil carrying an alternating current is brought in close proximity to an electrically conducting, non-ferromagnetic test specimen, and an alternating magnetic field is established, the alternating magnetic field causes currents, called eddy currents, to be induced in the conducting test specimen in accordance with Faraday’s law of electromagnetic induction. The alternating eddy current, in turn, establishes a field whose direction is opposite to that of the original or primary field. Consequently, the net flux linkage associated with the coil decreases. Since the inductance of a coil is defined as the flux linkage per ampere, the effective inductance of the coil decreases relative to its value in air. The presence of eddy currents in the test specimen also results in a resistive power loss. The effect of this power loss manifests in the form of a small increase in the effective resistance of the coil. An exaggerated view of the changes in the terminal characteristics of the coil is shown in Figure 1-3, where the variation in resistance and inductance is plotted in the impedance plane [2]. When a 1-2 Introduction flaw or inhomogeneity whose conductivity differs from that of the host specimen is present, the current distribution is altered. Consequently, the impedance of the coil changes relative to its value obtained with an unflawed specimen. Systems that are capable of monitoring the changes in impedance can, therefore, be used to detect flaws in a specimen that is scanned by a coil. The eddy currents exhibit a unique phenomenon known as the “skin effect” which causes the current density at a particular depth to decrease with an increase in the frequency of excitation. Skin depth (δ), also called standard depth of penetration, is defined as the depth at which eddy current density has decreased to 1/e of the surface value. The skin depth can be computed as follows: δ = 1 π∝f σ (1-1) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 where f is the excitation frequency of the circuit, μ is the magnetic permeability of the target material, and σ is the electrical conductivity of the target material. The skin depth is often used as a guideline to select the excitation frequency for a given test specimen. The variations in coil impedance caused by discontinuities in the test specimen are often very small in comparison with the background value of the coil impedance. The detection and measurement of the small changes is often accomplished using bridge circuits [2]. Alternating Current Eddy Current Conducting Specimen Figure 1-2 Eddy Current Generation and Flow in a Conducting Specimen 1-3 Introduction Figure 1 3 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-3 Impedance-plane Trajectory of a Coil over a Conducting Non-ferromagnetic Test Specimen [3] Factors that influence the eddy current field, and therefore the coil impedance, are as follows: The separation between the coil and specimen surface, called lift-off The electrical conductivity of the specimen The magnetic permeability of the specimen The frequency of the AC inducing the eddy current field The design of the eddy current probe Geometric factors Discontinuities, such as cracks, corrosion, pitting Successful detection and characterization of flaws requires a careful design of signal processing procedures to compensate for these effects. The elimination of undesired response and extraction and interpretation of relevant information forms the basis of considerable research activity in eddy current inspection. Multifrequency Eddy Current Techniques Single frequency eddy current tests offer excellent sensitivity to a number of different types of steam generator tubing under normal conditions. However conditions are often complicated by a number of factors and consequently inspection needs cannot be effectively solved by single frequency examinations. State of the art multifrequency eddy current testing overcomes most of the single frequency inspection limitations. The multifrequency technique consists of collecting data simultaneously using several excitation frequencies from just one probe pull. This provides data that are 1-4 Introduction analyzed using multifrequency mixing or multiparameter techniques. The technique not only allows the effect of extraneous discontinuities to be nullified but also improves the classification and characterization results. As mentioned earlier, each frequency is sensitive to a certain depth of the test sample. Low frequencies have large skin depths and hence generate strong indications of support structures that are located outside the tube. Thus, they are often used to determine location of support plates and other support structures along the tube. They can also be used to detect depositions of corrosion products on the outside of the tubes. Higher frequencies have a much smaller skin depth. There exists a well-defined relationship between the phase angle of the eddy current signal and the depth of defects, incorporated into the phase calibration curve, which is exploited to detect, classify and characterize the signals obtained from the high frequency channels. Due to the different skin depths at different frequencies (skin effect), signals from defects and support features change with frequency. In effect, this means that multifrequency response signals have additional information that can be analyzed to extract relevant features. The various advantages of multifrequency techniques can be summarized as follows [4]: http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Collects data at several test frequencies simultaneously. This decreases the in-service inspection time and human exposure time to radiation. Allows separation of discontinuities that give dissimilar signals at different frequencies. Improves sensitivity to different types of discontinuities. Improves the detection, interpretation and sizing of defects even in the presence of artifacts that complicate the analysis procedure. Two types of multifrequency probes have been predominantly used in recent years. The first is the bobbin coil probe, which consists of two identical coils connected in a differential mode and excited at multiple frequencies. The second type of probe is the rotating probe [5]. One benefit provided by this probe is to increase the resolution of tube degradation measurements. The bobbin coil probe is mainly used for the initial detection of possible degradation to quickly determine those areas of the tube requiring additional inspection with other types of probe that have improved ability to size and characterize degradation, such as rotating probes. Although the bobbin coil probe is the most widely used probe, it has limitations in its ability to detect degradation in all regions of the tube (e.g., expansion transitions), and hence these tube regions are further investigated by rotating probes. Other limitations of the bobbin coil probe include the ability to accurately size and characterize degradation. Bobbin Coil Probe A simple air core probe, when oriented coaxially in a tube, is called a bobbin coil. Figure 1-4 shows a differential bobbin arrangement, in which the signals from two identical bobbin coils are subtracted, in an attempt to provide a flaw signal that is more distinguishable from a relatively constant background signal [6]. When eddy current probes were first used for inspecting heat exchanger tubes, they were generally composed of two identical bobbin coils mounted closely together, operating in the 1-5 Introduction differential mode. A schematic of such a probe is shown in Figure 1-5. The recorded signals were the induced voltage in one coil subtracted from the voltage in the other coil. The advantage of this probe is that its signal is resistant to various anomalous effects, such as probe wobble, temperature variations, and gradual variations in the inspected tube’s electrical conductivity and diameter. This probe is very sensitive to abrupt anomalies, such as pitting corrosion and fretting wear [7]. Figure 1-6 shows a typical differential probe defect signal, measured in a laboratory, from a differential probe. The lower traces are strip chart displays of the respective imaginary (left) and real (right) components of the eddy current probe signals. Typical signals generated by a multifrequency-bobbin probe testing system are shown in Figure 1-7. The 35kHz low frequency channel is usually designed to locate the structure signal, such as tube support plate (TSP), top of tube sheets (TTS), etc. Research Objectives Although it is relatively easy to understand the basic eddy current probe-flaw interaction, realworld analysis of steam generator tube eddy current data is difficult largely due to noise and http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 many unwanted indications that cause significant distortion in the flaw signal. Automatic flaw detection systems for bobbin coil eddy current data have been well studied by many researchers [3,4]. These studies mainly focus on the relationship between flaw characteristics and the shape and orientation of the corresponding Lissajous pattern of the eddy current signal and attempt to mimic the decision process of a human expert. Figure 1-4 Differential Bobbin Probe, Comprising two Coaxial Air Core Probes in a Tube 1-6 Introduction http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-5 Schematic of a differential bobbin probe scanning a heat exchanger tube 1-7 Introduction http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-6 Typical differential probe signals. The lower traces are strip chart displays of the respective imaginary (left) and real (right) components of the eddy current probe signals. 1-8 Introduction http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-7 Multifrequency Bobbin Coil Probe Eddy Current Signals The objective of developing an automatic flaw detection system for analyzing steam generator eddy current data is to provide utilities with significant cost savings associated with reduced analyst requirements and faster inspections. Additionally, the added consistency and accuracy that automated data analysis potentially affords may allow utilities to demonstrate higher tube degradation detection probability and improved sizing accuracy. These capabilities could provide the basis for longer inspection intervals and the use of alternate repair criteria. Figure 1-8 shows the overall schematic of an automatic multifrequency bobbin coil probe eddy current data analysis system. The different modules of the system are described in detail in the next section. 1-9 Introduction Eddy current signals Preprocessing Classification Degradation No Degradation Determine Degradation Type Defect Characterization http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 1-8 Eddy Current Data Analysis System 1-10 2 SIGNAL PREPROCESSING The objective of signal preprocessing is to extract “meaningful” information from the data to be used subsequently for analysis. The raw signal is passed through a series of preprocessing routines that reduce the amount of data to be analyzed and classified by the classification algorithms. Figure 2-1 shows a schematic diagram of the overall preprocessing module. Raw Signal (from Multiview – AAPACK) Data Segmentation Ad ti Filt i http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Adaptive Filtering Dynamic Thresholding (Neyman-Pearson Detector) Moving Average Filter Distance Threshold Potential Indications Figure 2-1 Overall Structure of the Preprocessing Algorithm 2-1 Signal Preprocessing The various filtering operations carried out in the preprocessing routines as shown in Figure 2-2 are explained in detail in this chapter. Data Segmentation The objective of this step is to determine, in an automated fashion, the range of data points that need to be analyzed and those that need to be discarded (for instance, data points at the beginning – ‘air to tube entry point’ and end of the tube – ‘tube exit point to air’). Figure 2-2 shows the entry and exit points in the entire length of the tube data. The automated procedure records the maximum (near saturation) voltage reached at the exit point of the probe (transition from tube to air) and then scans for a voltage magnitude equal to 80% of the peak in the first half of the measurement data. If the algorithm does not detect any sample in the 80% range, then it picks the st 1 data point as “ the starting point within the tube ”, else the data point with voltage magnitude in the 80% range is chosen as “ the starting point within the tube ”. The exit point is chosen as the data point with maximum voltage peak at the tube end. http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 2-2 Entry and Exit point signals in the tube data 2-2 Signal Preprocessing Adaptive Filtering Adaptive filters are commonly used for minimizing time varying noise. Examples of its applications to nondestructive evaluation problems can be found in [8, 9]. In the case of steam generator inspection, such noise may be generated due to variations in liftoff due to probe wobble. An adaptive filter is capable of adjusting its impulse response appropriately using an algorithm that minimizes the error between the filter output and a reference input. We utilize a finite impulse response (FIR) filter whose coefficients are estimated using the least mean square (LMS) algorithm to implement the adaptive system. The overall adaptive filtering algorithm is implemented in two steps as depicted in Figure 2-3. The data is first passed through a normalized least mean square (NLMS) adaptive filter to remove time-varying noise from the data. This is followed by a wavelet based de-noising technique to remove any remaining random system noise. The reason for using a two-step procedure is due to the characteristics of the noise. Time varying noise requires an adaptive noise stochastic gradient descent cancellation scheme. Such an adaptive procedure typically requires a algorithm that tracks the stochastic properties of the noise in order to minimize it. The LMS and NLMS procedures are important members of this family of algorithms. The system noise, on the other hand, is not time varying. Popular digital filtering algorithms typically used to minimize non-time varying noise, such as low pass / high pass filtering usually designed on the basis of a trade-off between the sharpness of the filter characteristics and the window length. Shorter windows are required to minimize the computational effort, whereas wider windows are required for sharper filter characteristics [10]. Wavelet shrinkage denoising techniques provide an intelligent scheme that identifies the noise components and literally shrink them to reduce their effect on the signal. Such techniques do not use windowing schemes, and h f d ff f h i l i i f i l fil [11] http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 therefore do not suffer from the implementation issues of conventional filters [11]. The following sub-sections describe the processing steps associated with the proposed algorithm. Raw Eddy Current Signal NLMS Adaptive Filtering Wavelet Shrinkage Denoising Filtered EC Signal Figure 2-3 Block diagram of the adaptive filtering algorithm. Adaptive Filtering using the NLMS Algorithm Figure 2-4 shows the schematic of the adaptive LMS algorithm for noise cancellation. The idea underlying the approach is to exploit the correlation properties of noise in eddy current signals u k , and the and the signals due to defects and other artifacts in the tubes. The reference input, primary input, d k , to the adaptive system ideally are signals obtained from the eddy current probe containing only noise and from a probe passing over a defect respectively. In bobbin coil inspection, we have a choice of data at four different frequencies. Data from the higher frequencies, while containing correlated time varying noise, also contain flaw signals that can be 2-3 Signal Preprocessing correlated. Thus, one high-frequency signal (the 400 kHz signal is used as the reference in this study) is used as the reference signal, and the low frequency locator channel signal (35 kHz) is used as the signal containing only noise. We assume that the signal statistics change slowly. Signal Plus Noise dk Noise uk Adaptive Filter yk =? k – + Output Hk (z) εk Error Signal Figure 2-4 Schematic of the NLMS adaptive noise rejection system used for minimizing noise in eddy current data. In order to understand the function of the adaptive filter, consider Figure 2-4, along with the following equations: ds =+ kkk u kk =∋? ? (2-1) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 where d is the defect signal corrupted by noise, k u is the reference time-varying noise signal, k k sk is the defect signal itself, and ? k , ? ∋ represent the noise from the two channels. Let k k k y denote the predicted signal at the output of the adaptive filter. The underlying assumption is that the noise ? and ? ∋ , are highly correlated with each contained in the primary and the reference inputs, other, and uncorrelated with the signal component, k sk . When the adaptive filter coefficients are k optimized using the LMS algorithm, the algorithm converges to the minimum mean square error (MMSE) solution, y . This solution provides the best estimate, ? , of the noise contained in the primary input dk in the least square sense, that is, yk ? k . Since ε = s +? − y , the filter output k k k k is the target signal with a smaller noise component. The above argument shows that the adaptive d k and u k , which in this case is the filter allows cancellation of correlated components between time-varying noise component. Consequently, the error signal at the output of the noise rejection system provides an estimate of the desired defect signal component in the primary input signal. Since we have both horizontal and vertical components of both channels, we apply the adaptive filtering twice, once with the horizontal components, and once with the vertical components of the two channels. 2-4 Signal Preprocessing Wavelet Shrinkage De-Noising In the last processing step of the proposed algorithm, any residual system noise in the adaptive noise-cancellation system output is removed from the filtered eddy current data. This noise is treated as additive white Gaussian noise (AWGN), and a wavelet-based thresholding approach is utilized. The technique is known as adaptive wavelet shrinkage de-noising or soft thresholding w , of the eddy current data are "shrunk" towards [8]. In this method, the wavelet coefficients, zero using the relation, Γ?= ) sgn(w)[ w (w, −? ] + (2-2) The threshold, ? , depends upon the noise characteristics of the data and is estimated from the finest resolution level of wavelet transform of the data. Since the noise characteristics vary from probe to probe and from one tube to another, the threshold is computed adaptively for each tube. Zero-phase High Pass Filter The raw data usually contains unwanted low frequency components that can be removed using a zero phase high pass filter. In order to ensure a zero-phase, a Discrete Cosine Transform (DCT) is used for this purpose. The 1D discrete cosine transform is defined as (2-3) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 for The inverse DCT is defined as (2-4) for where in both (2-1) and (2-2) 2-5 Signal Preprocessing The DCT filter is applied to all the channels in the data. Figure 2-5 shows the effect of the DCT zero-phase filter on one of the channels. (a) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (b) Figure 2-5 (a) Original raw data, (b) High pass filtered (DCT) data 2-6 Signal Preprocessing Dynamic Thresholding (Neyman-Pearson Detector) This dynamic threshold algorithm makes use of the Neyman – Pearson detector that reduces the number of data points to be analyzed significantly. The adaptive (dynamic) thresholding algorithm calculates a variable threshold for consecutive segments of data and marks all signal points above the threshold as possible defect locations. The details of the Neyman – Pearson detector are provided in this section. The dynamic thresholding algorithm decomposes the mix-channel signal into smaller segments through a finite-length sliding window, and computes a threshold for each segment of data. The Neyman – Pearson (NP) detector is used to compute the variance-based optimal threshold that maximizes the probability of detection of an actual defect signal for a given probability of false alarm (PFA). Based on the thresholds, small segments of data points are marked as potential defect indications. Neyman-Pearson Detector This section explains the various steps of the NP detector algorithm with reference to the present application. 1. Background A Neyman-Pearson (NP) Detector maximizes the probability of detecting a signal (in presence of noise) for a given probability of false alarms (PFA). Basically, the detector is implemented by thresholding the output of a sliding window based test statistic. 2. Modeling Assumptions Let y(n), 0 = n = N-1, denote the adaptive filtered data samples, M denote the length of the sliding window, and X be the data vector comprising samples contained within the sliding i d http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 window. We assume that each sample of the high pass filtered data can be modeled as: X = W (Hypothesis H or, 0 ) 1 X = S + W (Hypothesis H ), where W is a non-defect signal (noise) and S is a defect signal. An NP detector defines the likelihood ratio, [| L () = PX H x PX H [| 1 0 ] ] (2-5) 2-7 Signal Preprocessing Let us assume that both S and W can be modeled as multivariate Gaussian signals, and in particular, suppose W = N (0,C W ) (2-6) S = N (μ S , C S ) where μ is the mean of S, and C and CS denote the covariance matrices of W and S, S W respectively. We further assume that noise is white, i.e.,is a diagonal matrix, W C CW = σ W I where σσσσ denotes the noise variance. W 2 2 (2-7) (2-8) 3. Neyman-Pearson Detector Under the above assumptions, taking the logarithm of the likelihood function, L(X), and simplifying the resulting expression, we can show that the NP detector becomes: Decide H 1 to be true if T (X) > δ, and Decide H 0 to be true if T (X) = δ , where T(X), known as the Test Statistic, is given by: Tx( )= − X T (C sw +σ 21I ) ∝s + 1σ 2 2 w X T [ 21 − C s (C s +σ w I )] X (2-9) and the threshold, δ, is determined from the desired probability of false alarms (PFA), i.e., http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 PFA = Probability { 4. Implementation Details T (X) > δ | H 0 } (2-10) Step 1: Approximate separation of signal and noise components A histogram of |y(n)| is used to approximately separate the two components. We compute a 1520 bin histogram of |y (n)| with outliers removed and identify the bins that dominate the histogram to be consisting mostly of noise points, with the rest being predominantly signal points. Call them w (n) and s (n), respectively. Step 2: Estimation of Sμ CS , σ W , 2 2-8 Signal Preprocessing Compute the variance of w (n) to get σ W . Subdivide s(n) into L segments, each of length M. Find the mean and covariance matrix of the “L” vector samples to Sget μ CS , respectively. and 2 (Note: If signal is assumed to be white, then simply compute the mean and variance of s (n) to get μS and σ S , respectively) 2 Step 3: Computation of the test statistic, T (X), for all data points Extend y (n) by padding (M-1)/2 zeros (assuming M to be an odd number) at the two ends. Use a sliding window of length M to compute T(X) at each point. Step 4: Determine PFA and choose a threshold, δ Suppose T W (X) denotes the vector comprising test statistic values of {w (n)} only. If PFA δ , as denotes the desired probability of false alarms, then compute the threshold, δ = (100 - PFA) th percentile of T W W (X). (Note: This means P {T (X) > δ} = PFA) Step 5: Selecting probable defect or non-defect indications Select all points where T (X)>δ as probable defect points, the rest being non-defects. The dynamic thresholding algorithm was applied to both the vertical and horizontal components of the mix channel to identify potential defect locations. The two sets of outputs (potential defect locations in the horizontal and vertical components) are combined to give the potential defect locations that need to be analyzed further. Figure 2-6 illustrates this process. Horizontal Component MIX NP Detector http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 MIX Vertical Component MIX Figure 2-6 Improved Dynamic Thresholding Algorithm NP Detector Potential Defect Locations Note that this algorithm may miss defects if the signal magnitudes are very small. These defects can be detected by appropriately adjusting the threshold parameters. However, this will result in an increased number of potential defect indications. Hence, the choice of the thresholds is a critical element of this algorithm. 2-9 Signal Preprocessing Figure 2-7 (a-b) and Figure 2-8 present the results of implementing this scheme. Original Signal 0.15 0.1 0.05 0 -0.05 -0.1 -0.15 -0.2 0 500 1000 1500 2000 2500 3000 3500 (a) 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 0 Possible defect locations marked by NP detection 500 1000 1500 2000 2500 3000 3500 (b) Figure 2-7 (a) Original raw data (Vertical, mix channel), (b) Output of Dynamic Thresholding (possible defect locations) Figure 2-8(a) shows a typical preprocessed signal and Figure 2-8(b) shows the zoomed in version of the specific defect. http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 2-10 Signal Preprocessing (a) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (b) Figure 2-8 (a) Illustration of the output potential defect points of the Dynamic Thresholding algorithm, (b) Zoomed in view of the defect region A fact to be emphasized is that any defect not marked by this preprocessing algorithm is not processed further and is hence missed. It is important to pick all potential defects at this initial 2-11 Signal Preprocessing stage, even though some of the potential defects may later be dismissed as a non-defect condition. Moving Average Filter The moving average filter is used to remove the high frequency noise riding on the potential defect signals in the vertical channel. Thus, this stage essentially acts as a low pass filter. The filter uses a 3-point window for smoothening purposes. If X(k) represents a vertical channel data point in the original data, and Y(k) represents the corresponding data point in the smoothed signal, then we have Y(k) = [X(k-1) + X(k) + X(k+1) ] / 3 Figure 2-9 shows the effect of the moving average filter on one of the channels. (2-11) (a) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (b) Figure 2-9 (a) Input to Moving Average filter, (b) Output of Moving Average Filter 2-12 Signal Preprocessing Distance Threshold For each signal section identified by the dynamic thresholding, the local minima in all channels is detected followed by the corresponding local maxima. Figure 2-10 illustrates how local minima are identified (from 200kHz channel) using the vertical channel information and verified using the impedance plane trajectory. The asterisks represent the actual defects and the circles represent the minima identified by the algorithm. (a) http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (b) Figure 2-10 (a) Locating the minima in the vertical component, (b) Corresponding minima represented in the impedance plane 2-13 Signal Preprocessing Further analysis of the minima-maxima pairs in the data allows us to establish a rule to distinguish between defects and non-defects as follows: Distance Rule: “If the distance between the local minimum and maximum in the vertical component (termed as the “ Min-max distance ) of the mix channel is greater than 3 data points, then this indication is treated as a potential defect, else it is considered as a non-defect signal.” This rule indicates that for defects and dents, the number of data points between the local minimum and the corresponding local maximum is more than 3 points. Results of the Preprocessing Module The performance of the adaptive filtering followed by the Neyman-Pearson Detector was analyzed on a tube-by-tube basis. The data distribution for each of the four plants in the EPRI OTSG bobbin database is shown in Table 2-1. The NP detector used a Probability of False Alarm (PFA) of 40%. Table 2-2 shows the final results of the preprocessing module. Table 2-1 Data distribution in the EPRI OTSG training database. Plant ANO CRYS OCN TMI # Defects (Expert Opinion) 42 49 30 34 # Dents (Expert Opinion) 4 5 - # Tubes 22 19 17 9 Table 2-2 Results of the Preprocessing Module Plant # Defects + dents before preprocessing 42+4 49 30 5 # Defects + dents after preprocessing 42+4 49 28 3 # Overcalls after preprocessing 11736 18913 14125 ANO CRYS OCN http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 OCN TMI Total 30+5 34 155+9 28+3 34 153+7 14125 8427 53201 2-14 Signal Preprocessing Table 2-2 indicates that, at the end of the preprocessing stage, we are left with 53361 potential indications. (Note: By “potential indication”, we mean the set of data points that constitute a single defect or dent). Thus, in all 53361 potential indications need to be further processed and classified. Out of the 155 defect indications, 153 defects were detected by the preprocessing stage. Two of the dents were missed at this stage. Table 2-3 shows the details of the missed indications after preprocessing. In addition to the two dents, the algorithm misses one MBM and one wear indication, all in the same plant. Table 2-3 Analysis of missed indications in the EPRI OTSG training database. Plant OCN OCN OCN OCN Cal. Group 27 56 27 27 Row 70 114 9 60 Tube 78 4 58 126 Indication 8469 7316 27281 15166 Category MBM Wear Dent Dent This classification algorithm is described in the next chapter in detail. http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 2-15 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 3 MULTISTAGE SIGNAL PROCESSING AND CLASSIFICATION A multistage approach of processing and classification was employed with the aim of reducing the number of overcalls in each stage while maintaining a high detection rate. Some of the key features that are used in these algorithms are the magnitude of signals and the calibration curve based phase-thresholds. The enormous number of potential indications obtained after the preprocessing stage is reduced using a multistage signal-processing algorithm as shown in Figure 3-1. A combination of rule bases and statistical classifiers are used to eliminate the non-defect indications systematically while retaining the defect and dent indications at each stage. In addition, certain classes of defects are handled separately. Volumetric indications, like wear and impingement, merit their own processing routines. These algorithms are applied to selected segments of data (usually at support plate locations) and any indications that are retained by these algorithms are merged with the indications remaining at the end of the main branch of the algorithm. The algorithms for wear and impingement detection are described at the end of this chapter. Magnitude Thresholds The magnitude thresholds are chosen based on a statistical analysis of the magnitudes of the potential defect indications in all the four channels. It is observed that most defects have a higher magnitude as compared to the non-defects. Thus, a relatively high threshold removes a large number of overcalls and yet picks up most of the defects. The magnitude thresholds are determined by using the statistical distributions of all the flaw and non-flaw indications in each plant. All the potential indications are then passed on to the “Calibration-curve based Phase Thresholding” stage. Calibration Curve based Phase Thresholds The phase calibration curves of the tubes can be used to adaptively compute the phase thresholds Figure 3-2 compares an ideal phase calibration curve with an experimental curve. Ideally, all defects have a phase angle in the range of 0 to 180 degrees. However, most of the shallow defects are generally too small for the eddy currents to generate a significant vertical component signal. Thus, in practice the calibration curves (especially, for the 200kHz and 400kHz channel http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 g , p ( p y, or low frequency channels) tend to have a range much less than 180 degrees. Figure 3-3 illustrates the calibration curves generated for one of the calibration tubes – R999C999G003. The phase thresholds computed for the four channels are indicated above each of the calibration curves. 3-1 Multistage Signal Processing and Classification Potential Indications Magnitude and Phase Thresholds Wear Rules Rule Base I Defects/ NDDs Rule Base II Impingement Rules Hidden Markov Models Dents Combine Non - Defects Defects Dents Figure 3-1 Block diagram of the Multistage classification module ID OD ID OD % Depth % Depth Phase Angle (ASME degrees) Phase Angle (ASME degrees) Ideal Calibration Curve Calibration Curve used in practice http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Figure 3-2 The ideal and practically used phase calibration curves 3-2 Multistage Signal Processing and Classification A few degrees are added to the phase thresholds computed from these calibration curves and the thresholds are applied to all the potential indications. The output of the phase thresholded data is then applied to the next set of rules to further reduce the number of potential indications. These rules are divided into two sections, and are designated as Rule Base I and Rule Base II in this report for simplicity. 200kHz - [3.4 91.6] MIX – [16.5 96.7] 400kHz – [17.8, 110.5] 600kHz - [4.0, 134.7] R999C999:Cal3 Figure 3-3 Phase calibration curves for all four channels of one of the calibration tubes R999C999G003 Rule Base I This stage makes use of the phase-trend feature of the defect indications. Phase trend refers to the variation of the phase angles of the indication with frequency. Defects are observed to have a http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 monotonically increasing or decreasing change in phase in the impedance plane trajectory (IPT) with frequency. This can be attributed to the skin effect phenomenon, described earlier. 3-3 Multistage Signal Processing and Classification Inner diameter (ID) defects tend to have a monotonic decreasing trend with an increase in frequency while outer diameter (OD) defect phase angles increase monotonically with frequency. This difference allows a classification of defects broadly into ID and OD flaw categories. In addition to the phase trend, this stage performs a second level of phase thresholding on the indications, where the thresholds are chosen for each frequency channel (200kHz, 400kHz, 600kHz and MIX) based on the available statistics of phase variation in the IPT. These thresholds differ for ID and OD indications due to the difference in the phase ranges of these two types of indications. In addition to defects, tubes in steam generators can contain dent indications that do not follow the phase trend. This stage also incorporates rules that screen out dents before performing ID-OD classification. The dent indications typically have a large peak-to-peak value in the horizontal component. Also the phase angles of these indications vary within a very small range close to the o o horizontal axis (i.e. ~0 or ~180 ). Both of these features are used to classify the dent indications. Figure 3-4 shows the IPT of a typical dent indication in all the four channels. Figure 3-4 The IPT of a DENT indication in all four channels http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 3-4 Multistage Signal Processing and Classification The phase angle of the indication is indicated by the lighter color line (red, if viewed in color) and is also indicated below the figure. The number adjacent to it, underlined by the dotted line, corresponds to the magnitude of the indication. Since ID indications also exist in the same phase range as dents, the rule-base checks for a phase trend on all the indications that are classified as dents at this stage. If the indication follows a trend, then it is classified as a potential defect. If the indication does not follow a trend, then it is classified as a dent. The indications that fail the dent rule are now passed through the OD-rule which checks for a o o monotonic increasing phase trend within a phase range of around 30 to 160 . This range is different for each frequency channel and is decided based on the phase ranges of the OD defects available in the training database. Figure 3-5 shows the IPT plots of an OD defect in all the four channels. This figure clearly illustrates the phase trend rule for OD flaws. Figure 3-5 The IPT of an OD defect in all four channels The indications that pass the OD-rule test are classified potential defects, and those that fail are applied to the ID-rule. The ID-rule looks for a reverse phase trend in the indications within a o o phase range of around 40 to -20 . Again, this range differs for each frequency channel. [Note: Phase angles in the 600 kHz channel for some IDs fall in the negative half of the impedance plane. This requires a negative bound for the phase range for ID.] Figure 3-6 shows the IPT of an http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 3-5 Multistage Signal Processing and Classification ID indication in all four channels and illustrates the monotonic decrease in phase angles with increasing frequency. Figure 3-6 The IPT of an ID indication in all four channels The indications that pass through the ID-rule are classified as potential defects and go to the next stage (Rule Base II), and those that fail the ID-rule are considered to be non-defect signals. Figure 3-7 summarizes the overall approach of Rule Base I. 3-6 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification RULE BASE I Dent P otential Indication Is it a D ent ? yes no ID I s it an O D ? I s it an ID ? yes no yes no OD no I s it an ID ? yes NDD Figure 3-7 Overall approach of Rule Base I. P otential D efects D ents Results of applying the magnitude and phase thresholds followed by the rules in Rule Base I on the four plants in the EPRI OTSG training database are shown in Table 3-1. The “# Defects + dents before Rule Base I” refers to the results following the preprocessing step. The “# Defects + dents after Rule Base I” refers to the results following the preprocessing step, the magnitude and phase threshold step and the Rule Base I step. Table 2-1 provides a summary of the distribution of defects and dents in the EPRI OTSG training database. After the Rule Base I step, 93% of the defects were detected (144 of the 155), Table 31. This represents a 90% POD at a 90% confidence level. A comparison in the number of defect overcalls + dent overcalls, before and after the Rule Base I step, shows an 88% reduction in the number of overcalls (53,201 before vs. 6,119 after). Table 3-2 summarizes the flaws missed in this stage. Note that a majority of the missed flaws are wear and impingement. A separate algorithm has been developed for detecting these flaws. Table 3-1 Summary of results after magnitude & phase thresholding, followed by Rule Base I. Plant # Defects + dents before Rule Base I 42+4 49 28+3 34 # Defects + dents after Rule Base I 40+2 48 22+3 34 # Defect overcalls after Rule Base I 1580 1725 1605 1048 # Dent overcalls after Rule Base I 64 40 23 34 ANO CRYS OCN TMI 3-7 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification Table 3-2 Summary of missed flaws/dents in Rule Base I. Plant ANO ANO ANO ANO CRYS OCN OCN OCN OCN OCN OCN Cal. Grp. 3 3 17 52 36 27 27 23 27 27 25 Row 44 44 150 31 46 108 108 71 108 137 135 Tube 48 47 11 40 37 111 111 126 111 71 11 Indication 24593 24576 3839 24552 1915 21570 20064 15666 15457 15856 12200 Category Dent Dent Groove IGA SCC (SP) ODIGA 1 st Stage Phase thrs. Phase thrs. Phase trend Phase trend Phase trend Magn. thrs Magn. thrs Phase trend Phase trend Phase trend Phase trend span Wear Wear Impingement Wear Wear ODI TSP Rule Base II All of the potential defect indications identified by Rule Base I are then processed through a second rule base (Rule Base II). Note that any indications that are labeled “Dent” in Rule Base I are not processed further. Rule Base II classifies the potential defect indications, from Rule Base I, on the basis of physical features. The data is classified based on the variance of the indications. The variance is computed for both the vertical and the horizontal component from all the channels. An analysis of the data indicated that the variance of the horizontal component does not contain any discriminatory information. However, a statistical analysis of the variance of the vertical component indicates that, by using an appropriate threshold, a large number of Rule Base I overcalls can be filtered out. This is indicated by the plot of the vertical variance for the 400 kHz channel (Figure 3-8). The red dots correspond to defects, as determined by expert opinion, and the blue dots correspond to Rule Base I overcalls. 3-8 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification Figure 3-8 Scatter plot of the variance of the 400 kHz channel for one plant. The threshold is selected on a plant-by-plant basis. The second physical feature is the cross correlation of the horizontal component and vertical component for each channel. In addition, the cross correlation across channels (for instance, the cross correlation between the horizontal components of the 200 kHz and 400 kHz channels) was also computed. Figure 3-9 plots the cross correlation between the horizontal components of 200 kHz and 400 kHz channels versus the vertical components of the same channels for one of the plants. The blue dots represent Rule Base I overcalls and the red dots represent defects, as determined by expert opinion. 3-9 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification Figure 3-9 Cross correlation between the horizontal components vs. the cross correlation between the vertical components (200 kHz & 400 kHz). The figure shows that defects lie in the extreme right hand top corner, suggesting a very high correlation. On the other hand, non-defects do not have a strong correlation. Additional information may also be obtained by using the correlation between the other channels. Results of applying Rule Base II to data from each of the four plants are shown in Table 3-3. Table 3-3 Results of Applying Rule Base II. # Flaws Input (Expert Opinion) ANO CRYS OCN TMI 40 48 22 34 # Flaws Detected by Rule Base II 38 48 20 33 # Overcalls by Rule Base II 370 375 361 511 Hidden Markov Models The Hidden Markov Model (HMM) [12] is a statistical model that has been successfully used to simplify the design and use of pattern classifiers. Statistical models are useful because they can account for the inherent variability in real-world systems. This provides the classifier a measure of robustness to measurement noise. The theory of HMMs was first developed by Baum et al 3-10 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification [13] and since then they have been applied to a multitude of areas [14, 15, 16]. Each signal is represented by means of a sequence of symbols, and HMMs are used to derive probabilistic models from the given data. All HMMs can be parameterized by means of a set of probability distributions that are derived from a training database. These distributions can be used to determine the likelihood that a given sequence of symbols was generated from the model. Since most real world signals are analog in nature, a clustering stage (called vector quantization [17]) is performed to convert the signal into a sequence of symbols. A typical sequence of steps in using a HMM would be as follows. A set of training sequences is obtained and used to generate the model. In a multiclass classification problem, we generate one model for each class. Additional data can then be used to refine the models and improve their capability. Finally, given an unknown sequence, we identify the model most likely to have generated the sequence. Eddy Current Classification The general scheme for the classification of eddy current signals is shown in Figure 3-10. A set of training signals is recorded and relevant features are extracted from each eddy current signal. Vector quantization of the data is carried out since discrete observation probability densities are used in the HMM. All four channels (200 kHz, 400 kHz, 600 kHz and Mix channels) are used, and the time sequence is replaced by a set of symbols. The set of vector quantized data are then used to train a HMM. Two models are developed, one each for defects and non-defects. During recognition, the unknown signal is vector quantized and applied to each of the models. The probability that the unknown signal was generated from each of the models is obtained, and the unknown signal is assigned to the class with the maximum probability. Results of applying the HMM algorithm to data from each of the four plants are presented in Table 3-4. Eddy Current signal Eddy Current signal Feature Extraction Feature extraction (Vector Quantization) (VQ) Test against stored models Select max. probability Recognition result Train HMM (a) Figure 3-10 (b) Flowchart for eddy current signal classification using HMMs (a) Training and (b) Testing 3-11 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification Table 3-4 Performance of the HMM on the EPRI OTSG training database. # Flaws Input to HMM ANO CRYS OCN TMI 38 48 20 33 # Flaws Detected by HMM 38 48 20 33 # Overcalls 94 95 72 100 Impingement Classifier A simple rule using only mix channel data was derived for classifying an indication in the TSP region as an Impingement or Non-impingement. Since, impingements are usually found in the TSP region, the algorithm only considers the TSP regions for impingement detection. The TSP region is divided into two halves, and each half is independently checked for impingements. The various features that are computed are as follows: Phase angle Magnitude V max −μιν ,H max−min V max − mean , V max− mean Figure 3-11 shows the vertical component and the IPT plot of the MIX channel of an impingement indication. 3-12 Multistage Signal Processing and Classification http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Multistage Signal Processing and Classification Entire TSP extent max V max-mean V max-min mean V min-mean min 1 st half 2 nd half Figure 3-11 Vertical component and IPT plot of an impingement As shown in the figure, most impingement indications have a prominent vertical component change above the mean (first in the negative direction, and then in the positive direction) in the region of the flaw. This feature is captured by computing the V values. These max-mean and V min-mean two are combined to form a normalized feature given by: norm V max min −− mean =⋅ VV max mean −− min mean VV min max −− ⋅ (3-1) max min The other feature of interest is the ratio of the Vertical peak-to-peak value and the Horizontal V max−min H max−min . peak-to-peak value and is expressed as Since the mix channel is being considered, one would expect to have close to a zero residue in the TSP region for non-defective indications. However, if an impingement exists in the region, the magnitude of the residue would be substantially higher. Thus, the magnitude was also used as a feature for the classification of the TSP indications. These 4 features were combined using Boolean rules for classification purposes. Table 3-5 shows the performance of this module on all the TSP regions of all four plants. 3-13 Multistage Signal Processing and Classification http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 Table 3-5 Results of the IMPINGEMENT CLASSIFIER on all four plants # Impingements (Expert Opinion) All Plants 8 # Impingements Detected (Auto Analysis) 8 # Impingement Overcalls (Auto Analysis) 129 Wear Identification Wear in steam generator tubes is a type of degradation that mostly appears at structure locations. Like impingements, these types of degradations are handled separately in the multistage classification algorithm. An analysis of the training data revealed that the residue in the mix channel at support plates has sufficient discriminatory information that can be used to differentiate supports with wear from ones without wear. Figure 3-12 shows the vertical and horizontal components of TSP signals without wear while Figure 3-13 shows the same components for support signals with wear. Comparison of the two figures reveals that the low frequency content in the residue of the vertical component is higher for support signals with wear. Thus, the energy in the low frequency coefficients of support signals can be used to discriminate between supports with and without wears. The energy is obtained by first computing the Fourier Transform of support signals and then computing the normalized energy in the low frequency coefficients (first 20 coefficients). Support signals from calibration groups that had wear were used to extract the features of interest (energy) and compute thresholds. This threshold was applied to all the support signals in a plant. 3-14 Multistage Signal Processing and Classification http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (a) (b) Figure 3-12 MIX channel vertical and horizontal components of two support plate signals without wear. 3-15 Multistage Signal Processing and Classification http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 (a) (b) Figure 3-13 MIX channel vertical and horizontal components of two support plate signals with wear. The results of applying the thresholds to support signals from the four plants are summarized in Table 3-6. 3-16 Multistage Signal Processing and Classification Table 3-6 Summary of the wear classifier. http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 # S S # TSP With Wear # Wear Detected # Wear Overcalls Plant ANO CRYS OCN TMI # TSP Signals 330 225 255 135 # TSP With Wear (Expert Opinion) 2 1 6 0 # Wear Detected (Auto Analysis) 0 1 5 0 # Wear Overcalls (Auto Analysis) 29 37 57 6 The results indicate that the wear classifier is capable of detecting most of the wear signals in the EPRI OTSG training database. However, the thresholds need to be optimally selected in order to improve the performance further. 3-17 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 4 SUMMARY AND CONCLUSIONS http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 A multistage automated classification algorithm for OTSG bobbin data has been described. An important point to note is that these algorithms are designed to be totally automated, with minimal operator input. Given a raw set of eddy current data, the algorithm undertakes a multitude of analysis procedures and identifies possible defect indications, including locations, without any intervention from an operator. The automatic data analysis system consists of two stages: signal preprocessing and degradation type classification. The algorithm can be summarized as follows: Preprocessing : This phase includes two major steps: adaptive filtering and dynamic thresholding. The adaptive filtering step removes correlated noise (such as noise due to probe wobble) by using tube-specific filters. The dynamic thresholding stage further removes low frequency trends and identifies potential flaw indications in the data. Classification : A multistage classification algorithm is used to classify potential flaw indications into different classes. The first stage consists of a rule base that classifies data into degradation signals (wear, impingement and NQI) and benign signals. The next stage (Rule Base II) reduces the number of NQI overcalls by applying a second set of rules to a combination of physical features. Finally, Hidden Markov Models are used to further reduce the number of NQI overcalls. The parameters for the rule bases and the HMM were modified to be plant-specific. A summary of the overall performance of the algorithm on the EPRI OTSG training database is shown in Table 4-1. This table includes the results of the wear and impingement classifiers. The results of these two classifiers have been combined with the results of the main branch of the classification algorithm, consisting of the rule base (Rule Base I and Rule Base II) and the HMM. As seen from the table, the newly developed algorithms detected 96% of the expert opinion defects (149 of 155) and achieved a 93% POD at a 90% confidence level, with 9.1 overcalls per tube. The missed flaws are listed in Table 4-2. Future algorithm development will focus on further reduction in the number of missed indications and a reduction in the number of overcalls. 4-1 Summary and Conclusions Table 4-1 Overall summary of the OTSG classification algorithm. Plant # Tubes # Defects + Dents (Expert Opinion) 42+4 # Defects + Dents Detected (Auto Analysis) 39+2 # Defect Overcalls (Auto Analysis) 151 # Defect Overcalls per Tube (Auto Analysis) 69 ANO 22 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 ANO CRYS OCN TMI Total 22 19 17 9 67 42+4 49 30+5 34 155+9 39+2 48 29+5 33 149+7 151 124 187 147 609 6.9 6.5 11.0 16.3 9.1 Table 4-2 Summary of missed flaws. Plant ANO ANO ANO CRYS OCN TMI Cal. Group 52 52 17 36 27 35 Row 31 31 150 46 70 45 Tube 40 40 11 37 78 16 Indication 24545 24553 9627 1917 8469 7085 Category SCC_SP SCC_SP Groove IGA Wear MBM ODI_TSP 4-2 5 REFERENCES http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 [1] W. Lord and R. Palanisamy, “Development of theoretical Models for NDT eddy current Phenomena”, Symposium on Eddy Current Characterization of Materials and Structures NBS, Gaithersburg, Maryland, September 1979. R. C. McMaster, P. McIntire and M. L. Mester, Nondestructive testing handbook, Vol. 4, 2nd Ed., American Society for Nondestructive Testing, Columbus, Ohio, 1968 Wiley Encyclopedia of S. S. Udpa, L. Udpa, “Eddy Current Nondestructive Evaluation”, Electrical and Electronics Engineering , edited by John G. Webster, Vol. 6, pp. 149-163, 1999 Electromagnetic methods of nondestructive testing , [2] [3] [4] , Edited by W. Lord, Gordon and Breach Science Publishers, London, 1985, pp. 175-304 [5] [6] [7] V. S. Cecco, Eddy Current Manual, Ontario, Chalk River National Laboratories, 1983. Kevin Kennedy Associates - http://www.kkai.com/mt12a.html The American Society for Nondestructive Testing http://www.asnt.org/publications/materialseval/solution/may00solution/may00sol.htm J. Kim, L. Udpa, S.S. Udpa, “Multi-stage adaptive noise cancellation for ultrasonic nondestructive evaluation,” Review of Progress in Quantitative Nondestructive Evaluation , vol. 18A, pp. 781-787, Plenum, New York, NY, 1999. Y. Zhu and J.P. Weight, “Ultrasonic nondestructive evaluation of highly scattering IEEE Transactions on Ultrasonics, material using adaptive filtering and detection,” Ferroelectrics and Frequency Control, vol. 41, no. 1, pp. 26-33, 1994. A.V. Oppenheim and R.W. Schafer, Hall: Englewood Cliffs, NJ, 1989. Discrete Time Signal Processing [8] [9] [10] , 2 ed., Prentice nd [11] D. L. Donoho, “De-noising by Soft-Thresholding,” Theory , vol. 41(3), pp. 613-627, 1995. IEEE Transactions on Information [12] L. R. Rabiner, “A tutorial in hidden Markov models and selected applications in speech recognition,” Proc. IEEE , Vol. 77, No. 2, pp 257-286, 1989. 5-1 References [13] L. E. Baum, T. Petrie, G. Soules and N. Weiss, “A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains,” Ann. Math. Stat., Vol. 41, No. 1, pp. 164-171, 1970. J. S. Bridle, “Stochastic models and template matching: Some important relationships between two apparently different techniques for automatic speech recognition,” Proc. Inst. Of Acoustics, Autumn Conf., pp. 1-8, 1984. R. P. Lippmann, E. A. Martin and D. B. Paul, “Multistyle training for robust isolated d h ii [14] [15] http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 word speech recognition,” Proc. ICASSP’87, pp. 705-708, 1987. [16] S. Zhong and J. Ghosh, “HMMs and coupled HMMs for multichannel EEG classification,” Proc. IJCNN’02, pp. 1154-1159, 2002. J. Makhoul, S. Roucos and H. Gish, “Vector quantization in speech coding,’ Proc. IEEE, Vol. 73, No. 11, pp. 1551-1588, 1985. [17] 5-2 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007 http://energysearch.epri.com/highlight/index.html?url=http%3A//www.epriweb.com/backup/download... 05/01/2007