EMD-based pulsed TIG welding process porosity defect detection and defect diagnosis using GA-SVM

EMD-based pulsed TIG welding process porosity defect detection and defect diagnosis using GA-SVM

Journal of Materials Processing Technology 239 (2017) 92–102 Contents lists available at ScienceDirect Journal of Materials Processing Technology jo...

3MB Sizes 1 Downloads 54 Views

Journal of Materials Processing Technology 239 (2017) 92–102

Contents lists available at ScienceDirect

Journal of Materials Processing Technology journal homepage: www.elsevier.com/locate/jmatprotec

EMD-based pulsed TIG welding process porosity defect detection and defect diagnosis using GA-SVM Yiming Huang a , Di Wu a , Zhifen Zhang b , Huabin Chen a,∗ , Shanben Chen a,∗ a b

School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an, Shaanxi Province 710049, China

a r t i c l e

i n f o

Article history: Received 2 March 2016 Received in revised form 26 June 2016 Accepted 10 July 2016 Available online 12 July 2016 Keywords: Arc spectrum EMD Feature extraction Hydrogen porosity GA-SVM

a b s t r a c t A portable spectrometer based on a linear CCD is designed with real-time acquisition and processing of spectral data in the welding process of aluminum alloys. An innovative method is introduced to diagnose and detect porosity defects. The method extracts several characteristic spectral lines and calculates the intensity ratio between H I and Ar I. The intensity ratio is used to diagnose extraordinary cases of hydrogen content. Empirical mode decomposition (EMD) is used to acquire adaptive decomposition of the ratio signal, which has been proved to have better performance in eliminating the influence of pulse current on the ratio signal than wavelet packet transform. Experiments based on X-ray inspection are designed to verify the proposed method. Monitoring of the arc atmosphere and detection of porosity under different welding processes is achieved by extracting the feature parameters. An improved support vector machine (SVM) classification model based on a genetic algorithm (GA) is built in order to guarantee accurate estimation of different types of porosity defects. © 2016 Elsevier B.V. All rights reserved.

1. Introduction One of the defects of the aluminum alloy welding process is porosity, which is prone to happen and causes huge damage to weld performance (Ascari et al., 2012). To ensure weld quality, nondestructive testing is always needed after welding. Nondestructive testing methods that can be used for detection of weld porosity include X-ray testing (Zou et al., 2015), ultrasonic testing (Yamamoto et al., 2014) and spectroscopy testing (Harooni et al., 2014). Valavanis and Kosmopoulos (2010) extracted 43 descriptors corresponding to texture measurements and geometrical features based on weld radiographs. The descriptors were used to identify the worm holes, porosity and other defects after training by SVM. A morphological image processing approach was proposed by Anand and Kumar (2006) to detect flaws, which were further categorized according to their properties from radiographic weld images. Sun et al. (2005) designed a set of real-time image acquisition systems based on X-ray to automatically detect defects such as slag inclusion, porosity and lack of penetration using a fuzzy pattern recognition method. Using an acoustic emission technique, weld-

∗ Corresponding authors. E-mail addresses: [email protected] (H. Chen), [email protected] (S. Chen). http://dx.doi.org/10.1016/j.jmatprotec.2016.07.015 0924-0136/© 2016 Elsevier B.V. All rights reserved.

ing porosity was detected in the structural component of a track crane with characteristic parameters such as amplitude and centroid frequency (Tao et al., 2014). Ultrasonic testing was employed by Kadumberi et al. (2012) to acquire data from a range of electrofused welding joints, and good corroboration between the observed weld quality and the ultrasonic data was achieved. However, both X-ray testing and ultrasonic testing have a higher requirement to the instruments and environmental conditions. In addition, it is difficult to distinguish porosity and slag inclusion using these test methods. Therefore, it is necessary to identify a novel method to detect porosity. An arc spectrum contains information about metal vapors, shielding gases and arc gases. Therefore, it is intrinsically linked to internal weld defects, which makes it the most promising new method for real-time detection of weld ˛ (2007) preliminarily studied the influence of defects. Weglowski welding parameters on the spectral intensity. A method was proposed to measure hydrogen content in the electric arc based on spectroscopy technology (Shea and Gardner, 1983). Sibillano et al. (2012) conducted welding experiments with three different types of laser source and found that there was a corresponding relationship between plasma electron temperature and penetration. Thus they put forward a method for on-line monitoring of the welding joint penetration. According to Kong et al. (2012), a correlation between the electron temperature and defects within the weld bead was identified. Mirapeix et al. (2007) used artificial neural

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

93

Fig. 2. Arc spectrum of aluminum alloy GTAW (peak current 240 A, base current 40 A). Fig. 1. Internal structure of the spectrometer: (a) Schematic diagram of the light path, (b) Spectrum sensor.

2. Experimental equipment and monitoring principle 2.1. Experimental equipment

networks to classify several types of surface defects by calculating the plasma electron temperature. In this paper, a defect detection system is developed based on spectroscopy technology that is aimed at solving the existing problems in the conventional detection method of porosity defects. The system achieves real-time processing of spectral data and warns of excessive hydrogen in the arc. Furthermore, a new method based on EMD is proposed for detecting weld porosity. The method is proven to be feasible and reliable through X-ray testing of welding beads. Finally, an SVM model based on GA was built to classify different types of porosity defined by ourselves and according to the international standard of classification of porosity.

Presently, commercial spectrometers on the market are used for general equipment, which means they are not designed for welding. Meanwhile, the secondary development of software for monitoring welding processes is costly. For these reasons, a portable spectrometer designed and manufactured for the welding process is discussed in this paper. The spectrometer achieves real-time collection of wavelengths ranging from 200 nm to 1100 nm in the aluminum alloy welding process. Moreover, a dimmer-filter glass system can be added to the internal structure to collect specific wavelength data as needed. The internal structure of the spectrometer is shown in Fig. 1. By acquiring arc radiation measurements in the welding process, a

Fig. 3. Diagram of the experimental system.

94

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

Fig. 4. The diagram of H atomic transition.

Fig. 5. Signals collected by data acquisition module (peak current 240 A, base current 40 A): (a) Characteristic spectral line of HI 656.27 nm, (b) Welding current, (c) Welding voltage.

Table 1 Components of 5A06 aluminum alloy and ER5183welding wire. Element

Mg

Mn

Fe

Si

Zn

Cu

Al

5A06 ER5183

5.8–6.8 4.3–5.2

0.5–0.8 0.5–1.0

≤0.4 0.4

≤0.4 0.4

≤0.2 0.25

≤0.1 0.1

Bal. Bal.

spectrogram is obtained that describes the function of wavelengths. The collected spectral data when the welding peak current is 240 A are represented in Fig. 2. The experimental system is shown in Fig. 3. 2.2. Monitoring principle for the extraordinary case of hydrogen content diagnosis During the welding process, hydrogen is absorbed and dissolved by molten aluminum alloy and precipitated from the

Table 2 Welding parameters. Group

Peak current

Base current

Welding speed

Wire speed

Thickness

Pretreatment

1 2

240 A 240 A

40 A 40 A

3 mm/s 3 mm/s

10 mm/s 10 mm/s

4 mm 4 mm

Remove oxidation film No pretreatment

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

greater probability of generating porosity when other parameters are constant. The key problem is determining the hydrogen content in the arc. When a hydrogen atom is charged by arc energy, it jumps from the ground state to an excited state. However, an atom in an excited state is very unstable. It will return to the lower energy state spontaneously. According to the selection rule for atomic spectra, the energy is released in the form of radiation, and atomic emission spectra are generated during a downward transition. The intensity of atomic emission spectra is expressed as (Kernahan and Pang, 1975):

Read signal s(t) Calculate the mean mi(t) of upper and lower envelope of s(t)

s(t)- mi(t)=hi(t)

hi(t) is used as new signal s(t)

Whether hi(t) meets the

No

Iij = Ni · Aij · Eij

conditions of the IMF

hi(t)is noted as ci(t) s(t)-ci(t)=ri(t)

Ni =

Yes

Whether

ri(t)

can

Z=

No c (t) + r (t)

molten aluminum alloy. Thus, there is a dynamic balance between the hydrogen in the molten pool ([H]) and that in the arc column (H2 (g)). According to Sievert’s law, the relationship between hydrogen content in the molten pool and the partial pressure of hydrogen in the arc is (Yu et al., 2015): 1/2 2

 (3)



 E  i

gi · exp −

kT

(4)

i

Fig. 6. Algorithm flow chart of EMD.

[H] = fH · S · PH



gi E · N · exp − i Z kT

Z is the partition function (Suzuki, 1976):

decompose into the IMF

s(t) =

(2)

where Ni is the population of an excited state atom, Aij is the transition probability between two states and Eij is the energy difference between two states. Because arc plasma is in a local thermodynamic equilibrium state, the hydrogen density distribution in the plasma, N, and the excited state atom density, Ni , follow the Boltzmann distribution law (Keefe et al., 1959):

Yes ri(t)is used as new signal s(t)

95

(1)

where fH is determined by the alloy composition and fH is the solubility of hydrogen in aluminum alloy at a given temperature. Under standard atmospheric pressure, the partial pressure of one type of mixed gas is proportional to its volume. Therefore, it is deduced that the greater hydrogen content in the arc is, the greater is hydrogen content dissolved in the molten pool, which leads to a

where gi is the statistical weight of the excited state, Ei is excitation energy, k is the Boltzmann constant and T is excitation temperature. Therefore, the spectral intensity is proportional to the total atom density. Under specific conditions, N is proportional to the amount of hydrogen. As a result, the spectral intensity can be used to characterize the amount of hydrogen. The choice of spectral lines also needs consideration. The process of hydrogen atomic transition is shown in Fig. 4. It is evident that the wavelengths of the resonance lines produced during the transition from the various excited states to the ground state are less than 122 nm. Wavelengths belonging to the ultraviolet range can be absorbed by nitrogen, oxygen, carbon dioxide and water in the air. The collection of these wavelengths is only feasible in a vacuum. When a hydrogen atom jumps from the second excited state to the first excited state, the emission line wavelength is 656.27 nm, which belongs to the visible light range and is easily acquired. In addition, HI 656.27 nm has the maximum relative intensity according to the NIST atomic spectra database. Therefore, HI 656.27 nm is selected as an important signal.

Fig. 7. The intensity ratio of spectral lines: (a) HI 656.27 nm and ArI 675.28 nm, (b) HI 656.27 nm and ArI 696.57 nm.

96

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

0.2

1 c1 c2-0.3 c3-0.7

0.5

c7-1.8 c8-2.1 c9-2.4

c4-1.1 c5-1.4 c6-1.6

c10-2.7 r10-3

0 -0.2

0

-0.4

-0.5

-0.6 -1

-0.8 -1.5

-1

-2

-1.2

-2.5 -3

-1.4 0

50

100

150

200

128

250

130

132

134

136

138

140

142

144

146

148

Welding length,mm

Welding length,mm

(a)

(b) 0.82

Reconstructed intensity ratio

0.8

0.78

0.76

0.74

0.72

0.7

0

50

100

150

200

250

Welding length,mm

(c) 1 c1 c2-0.4 c3-0.8

c7-2.1 c8-2.3 c9-2.5

0.34

c10-2.7 r10-2.9

0.33

Reconstructed intensity ratio

0.5

c4-1.3 c5-1.7 c6-1.9

0 -0.5 -1 -1.5 -2 -2.5 -3

0.32 0.31 0.3 0.29 0.28 0.27 0.26

0

50

100

150

200

250

Welding length,mm

(d)

0.25

0

50

100

150

200

250

Welding length,mm

(e)

Fig. 8. Decomposition and reconstruction of signals: (a) Components of IHI656.27nm /IArI675.28nm , (b) Partial enlarged detail of (a), (c) Reconstructed IHI656.27nm /IArI675.28nm , (d) Components of IHI656.27nm /IArI696.57nm , (e) Reconstructed IHI656.27nm /IArI696.57nm .

3. Design of experiments and data analysis 3.1. Experimental design As shown in Fig. 3, the experimental system consists of a weld module and a data acquisition module. The weld module includes a

robotic system, a water-cooled welding torch and a welder power supply (OTC INVERTER ELESON 500P). The data acquisition module obtains spectrum, voltage and current signals. The spectrum sampling period is 30 ms, while the voltage and current sampling rate is 500 Hz. 5A06 aluminum alloy plates with ER5183 welding wire are welded, and the components of each are shown in Table 1. In

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102 0.8

97

0.34

0.79

0.33

Denoised intensity ratio

Denoised intensity ratio

0.78 0.77 0.76 0.75 0.74 0.73

0.32 0.31 0.3 0.29

0.72

0.28

0.71 0.7

0

50

100

150

Welding length,mm

(a)

200

250

0.27

0

50

100

150

200

Welding length,mm

(b)

Fig. 9. Wavelet-packet denoise: (a) Denoised IHI656.27nm /IArI675.28nm , (b) Denoised IHI656.27nm /IArI696.57nm .

Fig. 10. Welding appearance: (a) Topside of weld beam, (b) Backside of weld beam, (c) X-ray testing result, (d) Longitudinal section.

Fig. 11. Reconstruction of signals: (a) Reconstructed IHI656.27nm /IArI675.28nm , (b) Reconstructed IHI656.27nm /IArI696.57nm .

250

98

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

Intensity,counts

Fig. 12. Welding appearance: (a) Topside of weld beam, (b) Backside of weld beam, (c) X-ray testing result, (d)Longitudinal section.

2000 1500 1000 500

0

50

100

150

200

Current

Welding length,mm 2 0 -2 1

1.05

1.1

1.15

Sampling

1.2 5

x 10

Fig. 13. Spectral intensity and welding current acquired from experiments in group 2.

order for exact control of the heat input and molten pool size, AC pulse GTAW is employed with a pulse frequency of 1 Hz and a duty ratio of 40%. Additionally, 99.99% argon with a flow rate of 14L/min is used as a shielding gas. A 3.2-mm-diameter tungsten electrode adding 2% ThO2 with a cone angle of 100 ◦ is employed. The welding process parameters are listed in Table 2.

3.2. Data analysis According to the previously stated monitoring principle, H spectral lines help to better understand the changing trends of hydrogen content in the electric arc. Due to the mutual aliasing effects of pulse current signals, H spectral lines appear to be non-linear and non-stationary, as shown in Fig. 5(a). Fig. 5(b) and (c) represent the welding current and voltage, respectively. Because the sampling frequencies of the spectrum and current differ, the range marked by red lines in Fig. 5(b) corresponds to the range marked by the rectangle in Fig. 5(a). Fig. 5 shows that the intensity of the spectral lines increases with an increase in the weld length. This is due to the accumulation of heat added to the ionization energy, which results in more hydrogen atoms. It is not clear whether the change in spectral line intensity is a result of the increase in the hydrogen content in the arc or from changing welding parameters. This affects the accuracy of weld porosity monitoring when using hydrogen spec-

trum information only. To solve this problem, the intensity ratio IH /IAr (H I spectral line with Ar I spectral line) is used to eliminate the effects resulting from factors other than hydrogen on the H I line. Ar I 675.28 nm and Ar I 696.57 nm are selected because wavelengths are close to H I 656.27 nm and have high relative intensities. The key to condition monitoring and defect detection is whether or not characteristics of the intensity ratio signal can be extracted accurately and comprehensively. This task can be effectively completed by empirical mode decomposition (EMD) (Huang et al., 1998). EMD is proposed to analyze data from non-stationary and non-linear processes. The major advantage of EMD is that the basis functions are derived from the signal itself. Hence, the analysis is adaptive, which is in contrast to traditional methods such as wavelet packet. Using a cubic spline interpolation algorithm, EMD self-adaptively decomposes the signal into multiple intrinsic mode functions (IMFs) with physical significance (Fig. 6). Any component of IMFs can satisfy the following criteria: (1) the number of extreme points equals or differs from the number of zero points by at most one; (2) the upper and lower envelopes are locally symmetric about the timeline. To eliminate the influence of pulse on the signal, the recovered signal is reconstructed with only a few low-frequency IMFs that are signal dominated and without high-frequency IMFs referring to pulse and noise. After analyzing the reconstructed signals, when the IHI656.27nm /IArI675.28nm ratio is greater than 0.74, the

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

99

Fig. 14. Spectral line of HI 656.27 nm: (a) Large and sparse porosity, (b) Small and dense porosity, (c) No porosity.

IHI656.27nm /IArI696.57nm ratio is greater than 0.3. From this, the conclusion is made that porosity exists in the weld seam. 4. Experimental results and discussion In the experiments in group 1, three spectral lines of H I 656.27 nm, Ar I 675.28 nm and Ar I 696.57 nm were extracted, and their IH /IAr signal intensity ratios were calculated, as shown in Fig. 7(a) and (b). The signal ratios in Fig. 7(a) were decomposed using EMD, and every component of the decomposition and the trend items are shown in Fig. 8(a). The curves in Fig. 8(a) show a tendency for the frequency to decrease from top to bottom. Fig. 8(b) shows partially enlarged details of Fig. 8(a). It can be seen that c1–c2 reflect noise, c3–c5 imply oscillations induced by pulse characteristics of welding source, and c6–c10 are low-frequency components where most of the signal structures are concentrated. The high-frequency components representing the pulse and noise were removed. The signal reconstruction is shown in Fig. 8(c). The

signal ratios in Fig. 7(b) underwent the same processing, and the results are shown in Fig. 8(d) and (e). The wavelet packet algorithm was used to reduce noise, and the results are shown in Fig. 9. Through analysis, it is determined that EMD retains more original signal features and eliminates the influence of pulse current. At the beginning of welding, the heat released by the electric arc continues to increase, hydrogen and argon increase differently in spectral intensity due to their differences in the degree of ionization. As a result, the IH /IAr signal intensity ratios have a noticeable deviation between the beginning of welding and the stable welding period. In this case, only the ratios associated with the stable welding period are analyzed. In the experiments of group 1 (Fig. 10), when there is no porosity defect, the IHI656.27nm /IArI675.28nm ratio is less than 0.74 and the IHI656.27nm /IArI696.57nm ratio is less than 0.3. Three spectral lines of H I 656.27 nm, Ar I 675.28 nm and Ar I 696.57 nm were extracted from the experiments in group 2 and their IH /IAr signal intensity ratios were calculated (Fig. 11). The welding beam in group 2 is shown in Fig. 12. It is found that when

100

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

Fig. 15. Statistical analysis results of three statuses: (a) Mean, (b) Standard deviation, (c) Kurtosis, (d) Skewness, (e) Range, (f) Confidence coefficient, (g) IHI656.27nm /IArI675.28nm , (h) IHI656.27nm /IArI696.57nm .

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

Begin Whether

Binary coding

meet

the

Yes

termination condition

No Fitness function of accuracy

Generate initial population

Calculate the fitness function

Fitness calibration

Selection operation Crossover operation Mutation operation

Determine the optimal solution, decoding Output the optimal solution

End

Fig. 16. Algorithm flow chart of optimizing parameters of SVM using GA.

there is a porosity defect, the IHI656.27nm /IArI675.28nm ratio is greater than 0.74 and the IHI656.27nm /IArI696.57nm ratio is greater than 0.3. When the welding beam was fully penetrated from 150 mm, the highest part of the energy in the arc moved downward. However, a fiber-optic probe used for collecting arc light was fixed, which led to a decrease in the spectral intensity collected by the spectrometer. The intensity of the H I spectral line decreased faster than that of the Ar I spectral line. As a result, the intensity ratio corresponding to a change in the penetration state decreased sharply. By analyzing the data acquired from the experiments in group 2, it was determined that a fluctuation in the spectral intensity was induced by variation in the hydrogen content. The fluctuation is irrelevant to the welding current, as shown in Fig. 13. Detailed extraction of the spectral data is seen in Fig. 14. To better distinguish the types of porosity, two forms of porosity were defined in detail based on the international standard of classification of porosity. The first type is large and sparse porosity, with a diameter between 0.2a and 0.4a (where a is the thickness of the plate in mm). The number of pores exceeds two, and the sum of the area is greater than 5% in a weld section of 2.5a × a (mm2 ). The second type of porosity is small and dense, with a diameter of approximately 0.05 + 0.2a. The number of pores exceeds eight, and the sum of the area is greater than 5% in a weld section of 2.5a × a (mm2 ). The intensity of H I 656.27 nm corresponding to different types of porosity was analyzed and is shown in Fig. 14. If the spectral intensity fluctuates widely in the period of peak value (fluctuation range greater than 150 and standard deviation greater than 60), it

101

implies that a large porosity is present. In the case of steady spectral intensity (fluctuation range lower than 100 and standard deviation lower than 40), there is no porosity. Small porosity is supposed to exist in the weld beam when the values of features fall between the above two cases. The segmentation marked in red in Fig. 14(a) indicates large and sparse porosity, yellow in Fig. 14(b) indicates small and dense porosity, and green in Fig. 14(c) corresponds to a sound weld with no porosity. Fifteen data points collected in the spectral line peak were used as a set of data, of which the mean, standard deviation, kurtosis, skewness, range and confidence coefficient were calculated. Forty sets of data were calculated and are shown in Fig. 15. It is observed from Fig. 15(a) to (f) that statistical characteristics of large porosity are quite different from the other two statuses. Moreover, the ratio signals in Fig. 15(g) and (h) can distinguish small porosity from no porosity. The eight feature values and three statuses above were taken as input and output factors of the support vector machine (SVM) model, respectively. SVM is a popular machine learning method for classification. The decision function classifying feature vectors is expressed as (You et al., 2015):

T

f (s) = sign(

t=1

¯ ˛t yt k (S (t) , S) + b)

(5)

where T is the number of training numbers, ˛t is a Lagrange multiplier, yt is a class label of training sample S (t) and b¯ is a bias. A radial basis function (RBF), k (S (t) , S), expressed in Eq. (6), is used to map samples into a higher-dimensional space (Jian and Gao, 2013):



k (S (t) , S) = exp −gS (t) − S2



(6)

where g is a kernel function parameter. The decision function was obtained by solving the dual optimization problem max˛

s.t.,



T t=1

T t=1

˛t −

1 T ˛i ˛j yi yj k (S (t) , S) 2 i,j=1

˛t yt = 0, 0 ≤ ˛t ≤ c



(7)

(8)

where c is a penalty factor. Two hundred sets of training samples were used to train support vector machines, and 40 sets were tested. By changing the kernel function parameter g and the penalty factor c, the accuracy of the SVM model varied from 80% to 90%. To eliminate the error resulting from random selection of relative parameters, a genetic algorithm (GA) was employed to optimize the parameters. The optimal parameters can be obtained due to the outstanding capability of global optimization of GA. A flow chart of this process is shown in Fig. 16. The prediction

Fig. 17. Classification results: (a) Fitness curve of GA, (b) Classification of GA-SVM.

102

Y. Huang et al. / Journal of Materials Processing Technology 239 (2017) 92–102

accuracy was taken as the fitness function of GA, and parameters corresponding to the highest accuracy in which the penalty factor is minimized were taken as the best parameters. It should be noted that the parameters first searched are selected as optimal parameters if several sets of parameters meet the condition that the penalty factor is a minimum. The evolutional generation was 200, the population size was 20, and the searching scope of the parameters ranged from 0 to 100. The best parameters obtained were a c value of 83.4316 and a g value of 0.61445. Finally, the classification accuracy reached 92.5%, as shown in Fig. 17. Label 1 indicates large and sparse porosity, Label 2 indicates small and dense porosity, and Label 3 indicates no porosity. 5. Conclusions This study aimed to construct a welding monitoring system for porosity defects of aluminum alloy AC pulse GTAW and then improve the weld quality and efficiency of the welding operation. The obtained results are summarized as follows: (1) A portable spectrometer was manufactured and equipped with a software system. This study has achieved real-time collection and processing of spectral data during the welding process, which provides device support for real-time control in the future. (2) The intensity ratio IH /IAr (H I spectral line with Ar I spectral line) is proposed to eliminate effects resulting from factors other than hydrogen on the intensity of H I line. This can be used in diagnosing hydrogen content, which is difficult. (3) Empirical mode decomposition is used to eliminate the influence of pulse current on the ratio signal. The presence of porosity is determined when the IHI656.27nm /IArI675.28nm ratio is greater than 0.74 and the IHI656.27nm /IArI696.57nm ratio is greater than 0.3. This conclusion is in agreement with the results of X-ray testing. (4) The intensity of spectral line H I 656.27 nm corresponding to different types of porosity was analyzed, and six statistical characteristic parameters were obtained. In combination with the two ratio signals above, an SVM model based on GA was built with a classification accuracy of up to 92.5%. Further studies concerning the detection of defects other than porosity are in progress. Acknowledgments This work is supported by the National Natural Science Foundation of China under Grant Nos. 51575349, 51275301 and national 863 plan under Grant No. 2015AA043102.

References Anand, R., Kumar, P., 2006. Flaw Detection in Radiographic Weld Images Using Morphological Approach, 39. NDT & E International, pp. 29–33. Ascari, A., Fortunato, A., Orazi, L., Campana, G., 2012. The influence of process parameters on porosity formation in hybrid LASER-GMA welding of AA 6082 aluminum alloy. Opt. Laser Technol., 44. Harooni, M., Carlson, B., Kovacevic, R., 2014. Detection of defects in laser welding of AZ31 B magnesium alloy in zero-gap lap joint configuration by a real-time spectroscopic analysis. Opt. Lasers Eng. 56, 54–66. Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.-C., Tung, C.C., Liu, H.H., 1998. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc. R. Soc. Lond. A Math. Phys. Eng. Sci. R. Soc., 903–995. Jian, L., Gao, C., 2013. Binary coding SVMs for the multiclass problem of blast furnace system. Industrial Electronics. IEEE Trans. 60, 3846–3856. Kadumberi, T., Prakash, A., O’Leary, R., Gachagan, A., 2012. Ultrasonic Imaging of Electrofusion Welded Polythene Pipes Employed in Utilities Industry. British Institute of Non-Destructive Testing, Northampton, pp. 129–139. Keefe, D., Nolan, P.J., Rich, T.A., 1959. Charge equilibrium in aerosols according to the Boltzmann law. Proc. R. Irish Acad. Sect. A Math. Phys. Sci. JSTOR, 27–45. Kernahan, J., Pang, P.-L., 1975. Experimental transition probabilities of ‘forbidden sulphur lines. Can. J. Phys. 53, 1114–1115. Kong, F., Ma, J., Carlson, B., Kovacevic, R., 2012. Real-time monitoring of laser welding of galvanized high strength steel in lap joint configuration. Opt. Laser Technol. 44, 2186–2196. Mirapeix, J., García-Allende, P., Cobo, A., Conde, O., López-Higuera, J., 2007. Real-time arc-welding defect detection and classification with principal component analysis and artificial neural networks. NDT E Int. 40, 315–323. Shea, J.E., Gardner, C., 1983. Spectroscopic measurement of hydrogen contamination in weld arc plasmas. J. Appl. Phys. 54, 4928–4938. Sibillano, T., Rizzi, D., Ancona, A., Saludes-Rodil, S., Nieto, J.R., Chmelíˇcková, H., ˇ H., 2012. Spectroscopic monitoring of penetration depth in CO2 Nd: Sebestová, YAG and fiber laser welding processes. J. Mater. Process. Technol. 212, 910–916. Sun, Y., Bai, P., Sun H.-y. Zhou, P., 2005. Real-time automatic detection of weld defects in steel pipe. NDT E Int. 38, 522–528. Suzuki, M., 1976. Relationship between d-dimensional quantal spin systems and (d + 1)-dimensional Ising systems equivalence, critical exponents and systematic approximants of the partition function and spin correlations. Progress Theor. Phys. 56, 1454–1469. Tao, Y., Wang, W., Sun, B., 2014. Nondestructive online detection of welding defects in track crane boom using acoustic emission technique. Adv. Mechan. Eng. 6, 505464. Valavanis, I., Kosmopoulos, D., 2010. Multiclass defect detection and classification in weld radiographic images using geometric and texture features. Expert Syst. Appl. 37, 7606–7614. ˛ Weglowski, M., 2007. Investigation on the arc light spectrum in GTA welding. J. Achievements Mat. Manuf. Eng. 20, 519–522. Yamamoto, S., Hoshi, T., Miura, T., Semboshi, J., Ochiai, M., Fujita, Y., Ogawa, T., Asai, S., 2014. Defect detection in thick weld structure using welding in-process laser ultrasonic testing system. Mater. Trans. 55, 998–1002. You, D., Gao, X., Katayama, S., 2015. WPD-PCA-based laser welding process monitoring and defects diagnosis by using FNN and SVM. Ind. Electron. IEEE Trans. 62, 628–636. Yu, H., Xu, Y., Song, J., Pu, J., Zhao, X., Yao, G., 2015. On-line monitor of hydrogen porosity based on arc spectral information in Al–Mg alloy pulsed gas tungsten arc welding. Opt. Laser Technol. 70, 30–38. Zou, Y., Du, D., Chang, B., Ji, L., Pan, J., 2015. Automatic weld defect detection method based on Kalman filtering for real-time radiographic inspection of spiral pipe. NDT E Int. 72, 1–9.