|NDT.net - July 2000, Vol. 5 No. 07|
|2nd International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, New Orleans May 2000.|
|TABLE OF CONTENTS|
| Ultrasonic Signal
||Preprocessing & |
| Feature Vector
|| Signal Class|
The acquired data is first subjected to a preprocessing step. Besides filtering for noise removal, this step also processes the signal for achieving invariance to selected inspection parameters. For instance, in the case of inspection data acquired at different inspection frequencies the signals are first transformed to an equivalent signal at a reference value of the inspection frequency parameter. Similarly, the overall classification performance of the system can be rendered invariant to other selected parameters . In the second step, discriminatory features in the signal are extracted. Feature extraction serves to reduce the length of the data vector by eliminating redundancy in the signal and compressing the relevant information into a feature vector of significantly lower dimension. The Discrete Wavelet Transform(DWT) is particularly effective at extracting features at multiple resolution levels in ultrasonic signals which are inherently non-stationary in nature. A second set of features based on Principal Component Analysis (PCA) also calculates the statistical properties of a set of neighboring A-scans [1,5]. The rationale underlying this approach is that the irregular nature of the IGSCC causes the variance of signals in the crack region to be vastly different from the variance of reflections obtained from a counterbore region. The PCA exploits this information to discriminate between cracks and counterbores.
Neural Network Classification
Neural networks are perhaps the most commonly used algorithm in automated classification of signals . These networks have proved to be extremely effective in learning subtle differences in signals from various classes (indications) as shown in Figure 2 . A neural network classifier with the error back-propagation training algorithm is used in the signal classification system developed in this project. The network was first trained using ultrasonic C-scan data from IGSCC samples generated using an automated scanner. A windows-based software package was developed for the analysis of C-scan image data. The system was tested using manually scanned signals from IGSCC samples.
System Validation on Fatigue Samples
The system had been previously trained on IGSCC and it was decided to determine its effectiveness in classifying fatigue cracks. The demonstrations were conducted using samples from the Performance Demonstration Initiative (PDI) program, a cooperative demonstration effort sponsored by utilities. The demonstrations were considered to be limited qualifications for austenitic and ferritic pipe. Small diameter piping and IGSCC were not included in the test set. Iowa State University (ISU) personnel operated the system at the EPRI NDE Center in Charlotte from May 10 to May 16, 1999. An employee from Duane Arnold Energy Center, Palo, Iowa, assisted in the data collection and reviewing the project. Alliant Energy Corp. has provided valuable support to this project. During this period data from PDI mechanical / thermal fatigue piping samples was collected.. An electromagnetic data acquisition and positioning system was used to collect ultrasonic waveforms manually from PDI samples. These were used to train the neural network. The Neural Network based ASC system was trained with this data to recognize the following classes:
The network was then tested with a different set of PDI test samples. After acquiring the data from the test samples the system was used to classify the indications. To truly test the neural network it was decided to not permit the operators to assist or confirm the classifications made by the network. The system was successful in detecting 100% of the cracks. Some of the crack lengths were measured from the processed data and they were accurate to within the error band allowed by PDI. More flaw lengths would need to be accurately sized to pass a PDI sizing test. The network had only one inaccurate signal classification. To maintain the sample security detailed results are not made available. The network was re-trained using the data from the false call. It was then tested on a similar sample and the network correctly classified the indication. A normal PDI demonstration would have required another test set to complete the demonstration.
Fig 2: Neural Network Classification Model
System Validation on IGSCC Samples
A great number of IGSCC examinations are performed using an automated system that would readily interface with a neural network for data analysis. It was decided to train and analyze signals obtained from previously recorded IGSCC database. The data was recorded in a proprietary format and log amplifiers were used to acquire the data which made it incompatible for neural network analysis.
Ultrasonic data collected with the automated system was sent to the researchers for neural network analysis. They were able to solve the two problems. The software has only analyzed linear data up until now. Using an intermediate file format (IFF) that was encouraged by EPRI several years ago and currently used by an ultrasonic equipment manufacturer, solved the first problem. The data was translated into the IFF prior to sending it to ISU researchers. The researchers developed an algorithm to handle the log data and display a C-scan image and A-scan data that is nearly identical to the original data, to solve the second problem. The benefit now available because of these solutions is that the NDEC and the industry have a large database of ultrasonic data that can now be analyzed by the neural network software. This will eliminate the need to acquire more IGSCC data from the practice samples. In-service UT examinations are currently performed in the field with the same techniques. The neural network could be used to analyze the field data. In the future the network could automatically analyze the data in parallel with the normal manual analysis. The data could be analyzed twice, similar to eddy current inspection of steam generator tubing.
Training, testing and validation with more IGSCC signals are in progress. Initial results of classification of the IFF data file are shown in Figure 3. The raw c-scan image of the file P-6A20.IFF is shown in Figure 3 (a) and the color coded classification image is shown in Figure3 (b) where the crack pixels are colored in red.
Fig 3: Raw c-scan image of the NDEC file P2-7A10.IFF(a) and the corresponding classification image with the IGSCC indicated in red.
|Figure 4. Display of (i) raw c-scan image from an IGSCC sample on the upper left hand window (ii) color-coded classification image on the upper left hand window (iii)A-scan signal at desired pixel in the lower left hand and (iv) Header information in the lower right hand window.|
Development of the Software Package for Commercial Use
The software package is in its final development stages. A sample screen display of the analysis software is presented in Figure 4.
|© NDT.net - firstname.lastname@example.org|||Top||