·Home ·Table of Contents ·Computer Processing and Simulation

The Investigation of Artificial Neural Network Pattern Recognition of Acoustic Emission Signals for Pressure Vessel Gongtian Shen, Qingru Duan, Bangxian Li and Qizhi Liu (National Center of Boiler and Pressure Vessel Inspection and Research) (China State Bureau of Quality and Technical Supervision) Contact

ABSTRACT:

The artificial neural network pattern recognition technique was employed to analyze AE sources signals of pressure vessel in site. A new quantitative analysis concept for AE sources of pressure vessel was introduced by using artificial neural network classification. A new method evaluating the severity of AE sources was raised. The designed and trained artificial neural network can give the percentage of crack growing, slag inclusion cracking, residual stress releasing and structure rubbing signals for an complex AE sources. This result let acoustic emission testing can non-destructively evaluate the safety condition of pressure vessel.

Keywords: pressure vessel, acoustic emission, artificial neural network, pattern recognition.

1. INTRODUCTION

Since Dunegan first applied acoustic emission (AE) technique (AET) to pressure vessel (PV) test in 1963[1], AET has been successfully used to a lot of pressure vessel test[2-4]. The final purpose of AE test for PV is to discover active defects. But many factors including defects and disturbing sources can produce rich AE signals during loading of a PV[5]. Defects include crack, incomplete penetration, incomplete fusion, slag inclusion, porosity and so on. Disturbing sources include residual stress releasing, oxide scales peeling off, mechanical impact, structural friction, leakage, electronic noise and etc. The key problem of AE test for PV is to distinguish AE signals produced by cracking and growth of real defects from large quantity of disturbing AE signals. Until now the natures of AE sources can not be distinguished by directly analyzing AE signals and must be verified by normal nondestructive testing methods such as visual test, magnetic particle test, dye penetration test, ultrasonic test and radiographic test.

However, it is very difficult to perform normal NDT verification for the fillet of nozzle, lad, skirt and support plates of saddles of PV. It is not able to perform normal NDT for on-line PV or PV with thermal insulation. These factors led can not make safety evaluation for PV. The use of AE test is restricted.

In recent years, a few papers report that the natures of some AE sources can be distinguished by artificial neural network pattern recognition of AE waveform[6-8].But most of AE instruments only can acquire and save AE characteristic parameters in the field pressure vessel test. This paper try to use the artificial neural network pattern recognition technique analyzing AE characteristic parameter to distinguish the characters of AE sources.

2. INVESTIGATION of BP NETWORK PATTERN RECOGNITION

Error back propagation (BP) neural network is one of neural networks most in common use. Figure 1 shows the structure of a 5X5X3 BP neural network with 11 input vectors and 3 output patterns. Figure 2 shows the principle of error back propagation. Through many times test, it is found that the best selection of input vectors of neural network for PV AE signals is following 11 parameters for one AE hit. The former 6 are the characteristic parameters of AE waveform. The later 5 are combined parameters from the characteristic parameters.
1. Rise time (R),     2. Count (C),    3. Energy (E),    4. Duration (D),     5. Amplitude (A).   
6. Count to peak (CP),    7. R/D,     8. C/D,    9. E/D,     10. CP/C,     11. A*R.

Fig 1: The structure of a 5X5X3 neural network

Fig 2: The principle of error back propagation

According to the requirement of AE test for field PV and the principle of convenient operating on personnel computer, a 50X50X5 BP neural network was designed. The 5 output patterns are classified as welding surface crack, welding inner crack, welding defects, residual stress releasing and mechanical hit or friction. The network was trained by 5 typical AE source signals that every one has approximate 500 AE hits. After 400 times training, the average squire error is 0.14 and the average correct recognition ratio is 93%. Table 1 lists the classification results of pattern recognition for training AE data and test AE data. The lowest correct recognition ratio is 86.5% and 84% respectively for training AE data of welding defects and test AE data of surface crack and welding defect. These results illustrate that the training effect of the network is very good. The trained network is possessed of very high generalizing capacity.

Output	Surface crack	Inner crack	Weld defect	Residual stress	Mechanical friction
Input
training data	89.0	97.5	86.5	98.1	99.5
test data
Mech.friction	0	0	0	0	100
Residual Stress	0	0	0	100	0
Surface crack	84	0	16	0	0
Inner crack	2.4	94.2	1.5	1.9	0
Welding defect	0	0	84	16	0
Table 1: The training and test results of network for 5 typical AE sources (Classification ratio %)

Table 2 lists the pattern recognition analyzing results performed by the trained network for 6 field PV AE sources listed in table 3. The recognition results of 5 AE sources in table 2 are basically in accordance with the normal NDT results in table 3. The total correct recognition ratio is 83%. Only the second AE source which is from welding scar surface crack is wrongly distinguished as mechanical friction. Through analysis the reasonable explanation was got as follows. Due to the biggest depth of the surface crack is 3mm and the thickness of the shell is 34mm, the surface crack can not grow under 2.0MPa pressure. But the AE data training network was produced by surface crack growing, so the patterns of these two AE sources is indeed different. In addition, the classification result of pattern recognition shows that the AE signals from welding scar surface cracks are produced by friction between crack faces.

Output	Surface crack	Inner crack	Welding defects	Residual stress	Mechanical friction
Input
No.1 AE source	9.8	82.5	5.4	2.3	0
No.2 AE source	0	0	0	0	100
No.3 AE source	10.8	0	83.8	5.4	0
No.4 AE source	0	0	0	0	100
No.5 AE source	0	0	25.2	13.0	61.8
No.6 AE source	0	0	39.6	25.5	34.5
Table 2: Recognition results of 6 field PV AE sources listed in table 3. (Classification ratio %)

No.	Description of AE sources	NDT verified results
1	Located on a longitudinal weld of a 1000m3 liquid petroleum gas (LPG) sphere.	One inner crack under surface 5mm with 15mm long and 10mm high.
2	Located on a welding scar of a 1000m3 LPG sphere.	3 surface cracks with respectively 15mm, 20mm and 30mm long, The highest depth is 3mm.
3	Located on a longitudinal weld of a 400m3 LPG sphere.	A lot of porosity, slag inclusion, incomplete fusion and incomplete penetration.
4	Spread on the fillet between shell and skirt of a heat exchanger.	No defect
5	Located on the fillet between shell and support lag of a 120m3 argon gas sphere.	A shallow surface crack with 10mm long and 0.5mm deep.
6	Located on a insulation support ring of a hydrogen cylinder.	The ring had been seriously corroded. There are a lot of oxides between ring and shell of cylinder.
Table 3: AE sources and NDT verified results of field PV

From table 2 we can find another significant result that the mechanism of AE sources can be quantitatively analyzed by artificial neural network. Due to training AE data of surface crack include AE signals produced by cracking of slag inclusion and the training AE data of welding defects include AE signals from residual stress releasing, the analysis performed by this network can not give accurate result of AE signals produced by different mechanism. In next section a new network was designed to analyze the mechanism producing AE signals.

3.INVESTIGATION OFRECOGNITION OF AE MECHANISM FOR PV

According to the analyzing results for AE sources of PV, except for leakage and electronic noise, all the other AE sources of PV are produced by the combination of crack growing, debonding and cracking of metallic oxides, residual stress releasing and mechanical friction. A new 50X50X4 BP network was designed with last 4 kinds of classification output. The network was trained by 4 typical AE source signals that every one has approximate 500 AE hits. After 300 times training, the average squire error is 0.02 and the average correct recognition ratio is 99%.

Table 4 lists the pattern recognition results of 8 AE sources of field PV performed by the trained network. The results show that artificial neural network pattern recognition can quantitatively analyze the mechanism of AE sources. The most significant finding is that the crack growing behavior of AE sources can be distinguished and the dangerous degree of AE sources can be quantitatively evaluated. For example, No.3 AE source is more dangerous than No. 8 although they are all welding defects. Even there are some surface cracks in No.2, but it is safety because the cracks did not grow. The recognition results of AE sources located on skirt fillet and lag fillet are satisfactory. The mechanisms produced AE signals include mechanical friction, oxides cracking and residual stress releasing.

Output	Crack growing	Oxides cracking and peeling off	Residual stress releasing	Mechanical friction
Input
1. Inner crack	94.5	0.5	4.3	0.7
2. Welding scar surface crack	0	97.7	0	2.3
3. Welding defect 1	17.9	2.8	76.2	3.1
4. Skirt fillet	0.1	23.2	0	76.7
5. Lag fillet	0	20.6	32.4	47
6. Support ring	0.1	20.1	76.1	3.7
7. Surface crack	57.2	1.0	9.5	32.3
8. Welding defect 2	0.6	2.8	96.1	0.5
Table 4: Recognition results of 8 field PV AE sources (Classification ratio %)

However, still there are some problems for the recognition results of this network. The AE signals produced by oxides cracking and residual stress releasing are easily confused. It is incorrect that 76% AE signals of insulation support ring was classified as residual stress releasing signals. The reason is that the training AE data were not produced by single mechanism.

4. CONCLUSION

1. Designed BP neural network can successfully distinguish the nature of field PV AE sources with the characteristic parameters of AE signals.
2. A new quantitative analysis concept for AE sources of pressure vessel was introduced by using artificial neural network classification. A new method evaluating the severity of AE sources was raised. The designed and trained artificial neural network can give the percentage of crack growing, slag inclusion cracking, residual stress releasing and structure rubbing signals for an complex AE sources.
3. The design of neural network structure and selection of training are very important factors affecting the recognition results. Using AE signals produced by as unitary as possible mechanism to train the network, the pattern recognition results will be better.

REFERENCES

1. H.Dunegan, "Acoustic Emission: A Promissing Technique", UCID-4643. Livermore, CA: Lawkence Radiation Laboratory (1963).
2. Shifeng Liu,etal,"AE inspection and defect assessment for metal pressure vessels", International Journal of Pressure Vessel and Piping, 38, P57-67, 1989.
3. T.J.Fowler, "Chemical Industry Applications of Acoustic Emission", Materials Evaluation, Vol.50, July, 1992: pp875-882.
4. R.D.Tidswell, M.P.Shipley and B.J.Cane, "Development of new procedures for in-service acoustic emission testing in pressure vessels and pipework", Progress in Acoustic Emission , The Japanese Society for NDI, Japan, 1996.
5. Gongtian Shen, Bangxian Li, Qingru Duan and Shifeng Liu: "Acoustic Emission Sources of Field Pressure Vessel Test", Progress in Acoustic Emission , The Japanese Society for NDI, Japan, 1996.
6. I.Grabec and W.Sachse,"Application of an Intelligent Signal Processing System to AE Analysis", Journal of Acoustical Society of America, Vol.85, No.3, 1989, pp1226-1235.
7. W. Schse and Igor Grabec, "Intelligent Processing of Acoustic Emission Signals", Journal of Materials Evaluation, Vol. 50, No.7, 1992: pp826-854.
8. R.S.Barga, M.A.Friesel and R.B.Melton,"Classification of acoustic emission wavesform for nondestructive evalution using neural network", in Applications of Artificial Neural Networks, Proc. SPIE, Vol.1294, pp545-556, 1990.

© AIPnD , created by NDT.net

|Home|    |Top|