NDT.net • Apr 2005 • Vol. 10 No.4

Development of an Automated Ultrasonic Testing System

Jiao Shuxiang
Robotics Research Centre, School of Mechanical & Production Engineering,
Nanyang Technological University, Singapore, 639798;
pg02502449@ntu.edu.sg

Stephen Brian Wong*
School of Mechanical & Production Engineering,
Nanyang Technological University, Singapore, 639798
mbwong@ntu.edu.sg; phone (65) 6790 4343; fax (65) 6791 4062.

Corresponding Author Contact:
Email: mbwong@ntu.edu.sg

© SPIE - The International Society of Optical Engineers.
This paper was originally published in the SPIE Proceedings vol. 5852
The 3rd Int' Conf' on Experimental Mechanics at 29.11.- 1.12.2004 in Singapore
Back To Session: Smart Structures and Non-Destructive Testing


ABSTRACT

Non-Destructive Testing is necessary in areas where defects in structures emerge over time due to wear and tear and structural integrity is necessary to maintain its usability. However, manual testing results in many limitations: high training cost, long training procedure, and worse, the inconsistent test results. A prime objective of this project is to develop an automatic Non-Destructive testing system for a shaft of the wheel axle of a railway carriage. Various methods, such as the neural network, pattern recognition methods and knowledge-based system are used for the artificial intelligence problem. In this paper, a statistical pattern recognition approach, Classification Tree is applied. Before feature selection, a thorough study on the ultrasonic signals produced was carried out. Based on the analysis of the ultrasonic signals, three signal processing methods were developed to enhance the ultrasonic signals: Cross-Correlation, Zero-Phase filter and Averaging. The target of this step is to reduce the noise and make the signal character more distinguishable. Four features: 1. The Auto Regressive Model Coefficients. 2. Standard Deviation. 3. Pearson Correlation 4. Dispersion Uniformity Degree are selected. And then a Classification Tree is created and applied to recognize the peak positions and amplitudes. Searching local maximum is carried out before feature computing. This procedure reduces much computation time in the real-time testing. Based on this algorithm, a software package called SOFRA was developed to recognize the peaks, calibrate automatically and test a simulated shaft automatically. The automatic calibration procedure and the automatic shaft testing procedure are developed.

Keywords: Ultrasonic nondestructive testing, pattern recognition, classification tree, ultrasonic signal processing

1. INTRODUCTION

Presently, most ultrasonic inspections are done manually. Most of the educational literature [1] on ultrasonic inspection has been written in technical language with engineering mathematics. Interpretation of ultrasonic NDT data can be a complex task, and interpreters must be highly skilled. When large engineering structures are inspected, the amount of data produced can be enormous, and a bottleneck can arise at the manual interpretation stage. Boredom and fatigue of operators can lead to unreliable and inconsistent results, where significant defects are not reported [2]. In order to improve their performance, it is recognized that there are gains to be derived from automating the ultrasonic inspection procedure by using a manipulator. Progress has been considerable in many aspects, namely, the development of equipment and techniques. For the automatic characterization of the ultrasonic transducer, M. S. Obaidant and D. S. Abu-Saymeh compared the performance of a neural network and statistical pattern recognition techniques [3]. The neural network method obtained a better result compared with the statistical pattern recognition techniques. An automated system based on a neural network for characterizing an ultrasonic transducer was proposed in 1998 [4].

Matthew Miller has developed an automated real-time data acquisition system for robotic weld quality monitoring [5]. K.M. Abd El-Ghany developed software that has the ability to use Internet to transmit testing data [6]. The expert can evaluate the ultrasonic scanning results for important structures without actually being present in the scanning site. Another useful tool is the simulation tool developed by Pierre Calmon [7]. The simulation tool can compute the ultrasonic fields of the realistic NDT configuration as an aid to NDT education and NDT test procedure optimization.

In spite of all the modern automated inspection methods, defect classification still is a difficult task. Much work, essentially based on pattern recognition techniques and artificial intelligence methods using neural networks, has been published on the subject.

Two methods, "best-fit analysis" and "triangulation analysis", were used to do shape classification [8]. Some success using neural networks has been reported with the use of the Hopfield [9] method and a back-propagation algorithm [10].

R. Drai [11] proposed a new signal processing technique known as "split spectrum processing". It improves the signal-to-noise ratio. The cross-correlation function is used to determine the resemblance between the echoes emanating from crack edges and small volumetric defects. "Split spectrum processing" was also used by Kwong Ki Yau [12].

I. Cornwel [13] used a rule-base expert system with fuzzy logic to deal with the problems of uncertainty during the automatic interpretation of ultrasonic NDT data [14]. There are also some other similar knowledge based systems [15]. Shaun W. Lawson describes the application of image processing and neural networks for the task of completely automating the decision making process involved in the interpretation of Time of Flight Diffraction (TOFD) images [16].

2. SIGNAL ANALYSIS AND ENHANCEMENT

Preprocessing is the first step after the data acquisition. To some extent, preprocessing decides the performance of the automatic system. It is necessary to carry out a thorough study on the ultrasonic signal before carrying on the project.

Fig. 1 is the partial signal from the ultrasonic card: SOFRA4001H using a 5MHz compression wave probe placed on a flat aluminum specimen of 25mm thickness.

Fig. 1. Partial signal.

The A/D converter on the card is 8 bits with a 40 MHz sample frequency. So the maximum amplitude is 255. The buffer for the signal is 8 K bytes. The digital signal is rectified. Rectification will blur the frequency character of the signal. This is one of the hardware limitations. Non-rectified signal cannot be got from the card, so all the analysis has to be based on the rectified signal.

1.1. Enhancement by cross-correlation
In digital signal processing, it is common to use cross-correlation to detect a signal or reduce noise [17].The premise of the cross-correlation includes:

  1. The noise is white noise.
  2. The sample set is large enough, the sample frequency is high enough.
  3. The target signal is known exactly.
However, the effect is not always so good for the peaks that are too high or too low. The main reason is that the template signal is not good for all the peaks. Another reason is that the data used to calculate the cross-correlation is too little.

1.2. Enhancement by zero-phase shift filter
Filters are very commonly used in the signal processing. Low Pass, High Pass, Band Pass, and Band Stop filters are all available, and the methods to design the filters are fully mature. Because the noise in the ultrasonic testing is mainly high frequency noise, low pass filter or band stop bank is desired. For this special application to reduce the noise for the non-destructive signal, it is necessary to design a zero-phase shift filter, because the peaks position should not been changed after filtering.

First the low pass filter is designed as:

A zero-phase shift filter based on the low pass filter is created.

1.3. Enhancement by averaging
Averaging is always the best way to eliminate the random white noise from the signal, when:

  1. There are enough data to do the averaging.
  2. The time to get enough data is acceptable.
  3. The time consumed by the averaging is acceptable.
The main advantage of averaging is that it will not distort the original signal definitely.

1.4. Enhancement result
The two figures (Fig.2 and Fig. 3) show the result of signal enhancement.

Fig. 2. Signal without enhancement. Fig. 3. Signal with enhancement.

3. AUTOMATIC DETECTION ALGORITHM

Whichever classification algorithm is adopted, the feature selected decides the performance of the classification algorithm. In many problems, extracting 'features' from the raw data will result in significantly improved pattern recognition. Features capture characteristics of the inputs that are most relevant for the estimation or classification problem at hand. Typically, domain-specific knowledge is used to develop good features.

Based on the signal analysis and the shape of the peaks, the selected features are:

  1. The AR Model Coefficients: 4-order AR Model is used to express the peak signals:
    Although more order for the AR model is possible for calculation, that will add more complexity and be not much effective because of the noise and the limitation of the sample data. The AR model coefficient a1, a2, a3 are selected as three features.
  2. Standard Deviation: The measure of dispersion or scatter of values of a random variable about the mean. If the value tends to be concentrated near the mean, the variance is small; while if the values tend to be distributed away from the mean, the variance is large.
  3. Pearson Correlation (Pearson's "r"): The measure of the linearity of the data.
  4. Dispersion Uniformity Degree: Dispersion Uniformity Degree is defined as:
    where: Maximum is the maximum value in the data set.
    ยต is the mean value of the data set.
    s is the standard deviation.

The next step is to choose the best set of features that discriminate the classes most efficiently. That is, enhance the separability among the different classes while increasing homogeneity within the respective class at the same time [18]. Because the signal is rectified, the frequency information got blurred. The separability shown by the AR coefficients is little. So the three AR coefficients are eliminated, only taking standard deviation s, Pearson's "r" and Dispersion Uniformity Degree are left as the features. This step reduces the computation complexity.

After the feature space is obtained, the next task of a pattern recognition system is classification. The goal of a classifier is to classify objects of interest into one of the number of categories or classes [19]. There are many methods to do the classification task. A neural network is the most often way mentioned in engineering papers. However, for the task to recognize the peaks, the classification tree is the better method.

A neural network classifier is suitable for the classification case shown in Fig. 3A. When the distance between class 1 and class 2 is short, in order to fulfill the classification task, more sample data is required to train the neural network, and of course more training time is also required to train the neural network [20][21]. However, for the classification task shown in Fig. 3B, no matter how many samples used, no matter how long the neural network is trained, it is not possible to fulfill the classification task perfectly for the common forward neural network [22][23]. The classification tree is not suitable for the classification task shown in Fig.3D, because the classification tree will produce a classification rule with zigzag form finally. That will result in a complex classification rule, or else a classifier with larger error rate. The classification tree is very suitable for the classification task shown in Fig. 3C, even the class-1 and class-2 have a small overlap part [24].

The task to classify the peaks in the non-destructive testing signal is much similar to the case shown in Fig. 4, just with a small overlap part.

Fig. 3. Comparing neural network and classification tree. Fig. 4. Peaks(red) and non-peaks(blue).

Fig. 5 is the whole tree generated according the splitting algorithm expatiated. The strings beside the nodes (triangle) are the condition to split the nodes. The leafs (dot) are the classified samples. "78" stands for the class "non-peak" and "80" stands for the class "peak". This is the whole tree, which means that all the data in the sample set can be classified successfully by the classification rule descripted.

Obviously, the rule is to complex and is hard to explain. And more, this is not the perfect classification rule because the sample set does not include the entire possible situation and there are error data that is misclassified. So pruning the tree is necessary. After pruning, the ultimate classification tree is shown in Fig. 6.

Fig. 5. Whole Classification Tree. Fig. 6. Ultimate Classification Tree.

It is easy to explain the classification rule shown in Fig. 6. That is: The samples are peak if they satisfy:

  1. The standard deviation larger than 2.49578
  2. And the dispersion uniformity degree larger than 3.20315 are peaks.
Or
  1. The standard deviation larger than 2.49578
  2. And the dispersion uniformity degree lower than 3.20315
  3. And the Pearson's "r" larger than 50.4576

4. RESULT

The table below is the experiment result. The automatic peak detection algorithm can detect the peaks with very little difference comparing with the human operator. When the gain is too high (higher than 45db), the signal will be much noisy. Even human operator cannot recognize all the peaks properly. In practice, almost all the test procedure defines the gain range.

Gain (db) Peaks recognized by eye Peaks recognized automatically error
6 2 2 0
10 2 2 0
15 3 3 0
20 4 4 0
25 4 4 0
30 9 9 0
35 16 15 1
40 20 21 1
Table: Experiment Result

Two automatic procedures, the automatic calibration procedure and the automatic shaft testing procedure, are designed and realized in the SOFRA software. The experiment result of the automatic calibration is good. Compared with the real defect position (Defect 1: 300mm, Defect 2: 505mm: Defect 3: 712mm), the experimental result of the testing on the aluminum shaft is very good (ref. Fig. 7). SOFRA also shows the capability to do the C-Scan. This shows the possibility to automate other ultrasonic non-destructive testing procedures

Fig. 7. Report Testing Result of the aluminum shaft.

5. CONCLUSION

Then, the statistical parameters are used as the features to do the classification task. For the real-time testing system, the outcome of using the classification tree to classify is much better than the outcome of using neural network to classify. The classification tree is used to decide the value of the features, and create the classification rule according to the ultimate optimized classification tree. The classification rule does the classification task very well.

REFERENCES

  1. Matthew J. Golis, ASNT Level 3 Study Guide Ultrasonic Method, The American Society for Nondestructive testing, Inc., ISBN 0-931403-29-4.
  2. Passi GS, Shoef Y, Kritsky MV, Reducing the influence of human factors on the reliability of manual ultrasonic weld inspection, Insight 1995; 37 (10); 788-791
  3. M. S. Obaidat, D. S. Abu-saymeh, Performance comparison of neural networks and pattern recognition techniques for classifying ultrasonic transducers, ACM 1992 0-89791-502-X/92/0002/1234
  4. M. S. Obaidat, H. Khalid, Balqies Sadoun, Ultrasonic transducer characterization by neural network, Journal of Information Sciences 107 (1998) 195-215
  5. Matthew Miller, Bao Mi, Akio Kita, I. Sharles Ume, Development of automated real-time data acquisition system for robotic weld quality monitoring, Mechatronics 12 (2002) 1259-1269
  6. K. M. Abd El-Ghany, M. M. Farag, New software for remote ultrasonic scanning via the Internet, NDT&E International, 35 (2002) 1-8
  7. Pieffe Calmon, Alain Lhemery, Isabelle Lecoeur-Taibi, Raphaele Raillon, Luc Paradis, Models for the computation f ultrasonic fields an their interaction with defects in realistic NDT configurations, Nulear Engineering and Design 180 (1998) 271-283
  8. McNab A, Dunlop I. AI techniques applied to the classification of welding defects form automated NDT data, Br J Nondestructive Testing 1991;33;11-18
  9. Baker AR, Windsor CG, The classification of defects form ultrasonic data using neural networks: the Hopfield method, NDT Int 1989;22(1);97-105
  10. Santos JB, Perdigao J, Automatic defects classification system using spectral analysis and neural networks, Ultrasonic International 93 Conference Proceedings, 1993; p763-766
  11. R. Drai, F. Sellidj, M. Khelil, A. Benchaala, Elaboration of some signal processing algorithms in ultrasonic techniques: application to materials NDT, Ultrasonics 38 (2000) 503-507
  12. Kwong Ki Yau, Split-Spectrum Processing for Nondestructive Testing, NDTnet August 1997, Vol.2 No.08 www.ndt.net/article/splitspec/splitspec.htm#fig26
  13. Jack Minker, Logic-based artificial intelligence, KLUWER ACADEMIC PUBLISHERS, ISBN 0-7923-7224
  14. I. Cornwell, A. McNab, Towards automated interpretation of ultrasonic NDT data, NDT&E International 32 (1999) 101-107
  15. Jarmulak J., Kerckhoffs E. J. H., van't Veen P. P., Hybrid knowledge based system for automatic classification of B-scan images from ultrasonic rail inspection, Tenth Conference on Innovative Applications of Artificial Intelligence, 1998 p1121-1126
  16. Shaun W. Lawson, Graham A. Parker, Automatic detection of defects in industrial ultrasound images using a neural network, Proceedings of SPIE, vol 2786, 1996, pp. 37-47.
  17. Professor Lordon S. Kino, Acoustic Waves: Devices, Imaging, and Analog Signal Processing, Prentice-Hall, Inc. ISBN 0-13-003047-3
  18. Sankar K. Pal, Amita Pal, Pattern Recognition from Classical to Modern Approaches, World Scientific, ISBN 981-02-4684-6
  19. Charles W. Therrien, Decision Estimation and Classification, By John Wiley & Sons, ISBN 0-471-83102-6 [44]. Menahem Friedman and Abraham Kandel, Introduction to Pattern Recognition-Statistical, Structural, Neural And Fuzzy Logic Approaches, World Scientific, ISBN 9810233124.
  20. B.D.Ripley , Pattern Recognition and Neural Networks, Cambridge University Press, ISBN 0 521 46086 7.
  21. Martin Anthony and Peter L. Bartlett, Neureal Network Learning: Theoretical Foundations, Cambridage University Press. ISBN 0 521 57353.
  22. STAMATIOS V. KARTALOPOULOS, Understanding Neural Networks and FuzzyLogic, IEEE Press, ISBN 0-7803-1128,1996.
  23. Simon Haykin, Nerual Networks: A Comprehensive Foundation, Macmillan College Publishing Company, Inc., ISBN 0-02-352761-7.
  24. Ruby L. Kennedy, Yuchun Lee, Benjamin Van Roy, Christopher D. Reed, Dr. Richard P. Lippmann, Solving Data Mining problems through Pattern Recognition, Prentice Hall, ISBN 0-13-095083-1
© NDT.net |Top|