Table of Contents ECNDT '98
Session: Nuclear Industry
Neural Network Classification of Flaws Detected During Ultrasonic Inspection of Candu Pressure Tubes and Brazed Joints in Candu FuelRazvan Ciocan, Dorin Ciobanu
Institute for Nuclear Research - Pitesti, Romania*
Wade R. Mayo
Chalk River Laboratories, AECL Research, Canada
|TABLE OF CONTENTS|
The use of Neural Network (NN) software packages has two attractive features. These are: rapid design and use of a certain NN type, and finding the optimum parameters in the neural network learning process. The simulation algorithms on PC computers can be implemented using a conventional programming language.
Results obtained by using a NN for defect classification of ultrasonic signals have been reported in the technical literature. Karpur  demonstrated the possibility of using a NN for detection and classification of discontinuities in composite materials. The radio-frequency (RF) signals were acquired, and a Fast Fourier Transform (FFT) was performed on them. The NN thus designed (based on a back-propagation algorithm) used the A-scan data and spectral amplitude of each acquired signal. With this NN, the detection probability for a discontinuity was 71.11%, and the probability of obtaining a false result (i.e., signaling of a discontinuity when one is not present) was 15.58%. Other authors have reported on the use of data obtained by ultrasonic spectral methods in neural network teaching patterns. A detailed presentation of this procedure is found in paper .
The essence of the spectral method is that the flaws are irradiated by ultrasonic oscillations in a wide band of frequencies. The oscillations are scattered from the defect in the direction of the search transducer, the echo signal obtained is selected in time, and the transfer function of the defect is analyzed. The spectrum of the echo signals reflected from flaws depends on the shape of the flaw via directivity functions. The authors in paper  have used the following parameters of ultrasonic signal spectra: spectrum amplitude, peak frequency, and bandwidth, in order to identify and classify an artificial flaw with a neural network.
This paper presents the results obtained with a NN for classification of defects. The input data used in the learning stage were the radio-frequency signal and its Fourier Transform, obtained from artificial flaws. This NN was tested for two situations : artificial flaw in a (Zr-2.5%Nb) CANDU pressure tube, and defects in the brazed joints of CANDU nuclear fuel.
The nondestructive examination of CANDU pressure tubes for flaw detection is performed to detect two broad types of flaws. These are flaws that originate during manufacture of the tube, and those induced during in-reactor service. The standard ultrasonic methods of detection are fundamentally the same for both types. Two opposed axially and two opposed circumferentially-directed shear wave beams, and a normal beam interrogate the tube . Once detected, the flaws that occur during manufacture are subjected to a threshold test. If their signal amplitude exceeds the signal from a standard reflector, the tube is reworked or rejected. The service induced flaws receive more characterization effort, since replacement in an operating reactor is expensive in terms of economic cost and radiation exposure.
Flaws that occur during pressure tube manufacture usually result from discontinuities that can be traced back to the production of the ingot from which several, typical twenty to thirty, tubes are made . These discontinuities are deformed by the same processes that produce the thin-walled tube from the massive ingot, namely forging, extrusion, and cold drawing. They therefore usually appear as axially-elongated planar discontinuities, are relatively tight, and often are filled with metallic oxides. They usually break the inside or outside surface, although one or two exceptions to this rule have been seen, and they are often angled with respect to the tube surface. On rare occasions, pits have been seen in finished tubes after manufacture.
Service induced flaws include mechanical damage caused by fuel movement, debris frets, and the effects of operation on flaws produced at the manufacturing stage that were not rejected. Mechanical damage caused by fuel movement results from particulate matter that may find itself lodged under fuel bundles. During the refueling operation, the particulate may gouge the inside tube surface when the bundles are pushed. This is gouging of the tube over and above scoring by fuel bundle bearing pads, and can result in minor flaws that are detectable and require characterization. Metal curls and proud-of-the-surface lips often accompany them. Debris fretting results from particulate that circulates in eddies in the fuel channel and scours out pockets in the tube surface . These are usually fairly broad and pit-like. The third type of service induced flaws, those that result from effects of service on un-rejected flaws that occur during manufacture, are of particular concern. Some flaws that occur during manufacture can be of sufficiently low signal amplitude that they will not be rejected. After some ten years of operation under conditions of temperature and pressure cycling, and irradiation, they may grow to a state where they are defects. When this happens, the tube must be removed. The frequency of occurrence of this phenomenon is extremely low, and flaws that grow to the extent of becoming defects are easily detected, if the tube wherein they lie is inspected.
It should be noted that the nondestructive examination of pressure tubes for flaw detection also involves the characterization of indications that are not indicative of flaws. For ultrasonic examination, these can include such items as grooves and flat spots, outside surface installation scratches, and the scoring by fuel bundle bearing pads alluded to above, known as "fueling tracks." All these phenomena produce signals during the detection phase of in-service pressure tube examination, which must be correctly characterized as benign. The majority of pressure tube indications usually fall in this category.
The measurements on pressure tube material were performed at Chalk River Laboratories (CRL) Nondestructive Testing Development Branch. Six axial EDM notches, with tip radii ranging from 0.1 mm to 1.5 mm, were scanned. The notches were identical in depth (1.5 mm) and length (12 mm) and were machined on a 0.46 metre long (Zr-2.5%Nb) pressure tube stub. An ultrasonic transducer of 33 mm focal distance and 15 MHz nominal frequency was used, placed outside and aimed normal to the pressure tube outside surface. Each notch was scanned over two circumferential 45° arcs. Data acquisition was performed using a software package named General Scanning Interface Library (GenSIL) which was developed at CRL for collecting ultrasonic, eddy current and acoustic emission data. This program performs the acquisition, both saving and showing the data in near real time.
Brazed joints were investigated at the Institute for Nuclear Research (INR) Pitesti, with a high frequency transducer (50 MHz nominal frequency). The program used for ultrasonic investigation of the brazed joint was ULTRAPAC II . The principle of ultrasonic investigation of brazed joints is described in detail in paper . The processing of data (i.e., obtaining FFT from the acquired data) was by a special package designed for this application, with the same program used to process data obtained from both pressure tube and brazed joints. The artificial flaws used in the learning process of RNBRAZ (the NN designed to recognize the defects in brazed joints) were those used as calibration standards for the ultrasonic investigation. They were:
Coding of the NN simulator was in Visual Basic V4.0. The program was developed on an IBM PC AT-386 computer with 4MB RAM and 33MHz clock rate. This program can be used on any IBM-PC compatible computer with Windows 3.1 installed, or any comparable system.
The results presented in this paper are based on the back-propagation algorithm. Due to the design of our NN facilitates, including updating and the other algorithm types, the number of levels was limited to 7 by using a recursive algorithm . This limit avoided problems associated with exceeding storage capacity, and was used only for testing purposes. (The learning process is influenced by many hidden layers.) There is no limit to the number of neurons allowed on each level, but the available RAM limits the total number of neurons by limiting the number of mathematical operations.
The following transformation functions, used in the NN are indicated in :
The factor "t" is called neuron temperature, and contributes to the convergence of the learning process. Our NN also permits the use of 3 other functions. These functions have been obtained as a combination of the first three, having a strong non-linearity around zero:
Our NN can use one of the functions F1 to F6 on each level, except the input level; each level has the same transformation function for all neurons on that particular level. The input patterns have the same structure for each problem. The output is dependent on the number of defect classes considered. Therefore we trained two NN's:
RNPT has seven defect classes as output:
The following defect classes were used as outputs for RNBRAZ:
The signal obtained from the artificial flaws was used in the training process for each NN. The amount of data was different for each NN and for each defect class. In preparation for testing, the signals obtained on artificial flaws were used. These were obtained from a different scan than that performed for learning process signals. In order to test RNBRAZ, a signal acquired from a natural flaw detected in a brazed joint was used. Under such conditions, we consider the data used in the process of testing and in learning to be different, having been obtained from independent measurements.
The results obtained in the process of testing of RNPT are presented in Table 1. We conducted 12 tests with this NN. The best results were obtained for classes G3 and G4 (0.8 mm and 0.6 mm tip radii). The detection probability in these cases were 0.999533 (T5) and 0.999533 (T6). Good reproducibility in the classification process was obtained for flaw G5. The values obtained were: 0.886653 (T3) and 0.885023 (T4).
We consider RNPT to also have obtained acceptable results for flaw G6: 0.800791(T1) and 0.869193 (T2). Note that in results obtained for test T2, the probability of classifying G6 in the G5 flaw class is 0.257822. Good results were also obtained in T11 and T12 tests for the G0 (flawless) class: 0.924688 and 0.910881 respectively. The amount of data used in the learning process affects the results obtained in the testing stage. This can be seen from results obtained for tests T7 to T10. The data for these tests were obtained from flaws G1 and G2 ( 1.5 mm and 1 mm tip radii, respectively). The data obtained from the second scanning of G1 were not available, thus the amount of data used in the learning process for the G1 class was half that used for the other flaw classes. The signal used for T10 was that obtained at 2° from normal at the tip. For these reasons, the results are contradictory for this test: 0.405527 detection probability for G1 and 0.539297 for G2.
Only two tests were completed for RNBRAZ, because a greater number of patterns was used in the training stage. This NN was tested with the signals obtained from the natural flaw. The results obtained are presented in Table 2.
The main purpose of the testing process of the RNBRAZ was to determine the probability of detection for a natural flaw in a brazed joint. The signal with the greatest amplitude (50% of that used in calibration) from this defect was used in the testing process (T1). The results obtained with RNBRAZ place this defect with a 0.991564 probability in the C1 class, and 0.299178 probability in the C3 class. The results obtained could not be confirmed by the ultrasonic image, but a succession of metalographic sections confirmed the RNBRAZ results. The defect was a pore-chain, with the longest dimension along the axis of the brazed joint. The test T2 was performed on an artificial flaw from the C3 class. The NN classified this defect undoubtedly: 0.998441 detection probability.
The work presented in this paper had two goals:
The results obtained in this paper show that a neural network could be used for fast classification of real defects detected by ultrasonic examination. The actual NN could be developed, and training could include the natural flaw, previously detected, during the learning stage. The present system for acquisition and "intelligent" analysis of ultrasonic data represents an important step in the realization of the Intelligent System for Ultrasonics (ISU) proposed in paper . This system could be used not only in the fast identification and classification of defect, but also in the integrity characterization of materials.
The authors wish to thank Dr. A. Wei, formerly of AECL, for his assistance in collecting the ultrasonic data at CRL