Investigation on Ti6Al4V laser welding using statistical and Taguchi approaches

Investigation on Ti6Al4V laser welding using statistical and Taguchi approaches

Journal of Materials Processing Technology 167 (2005) 422–428 Investigation on Ti6Al4V laser welding using statistical and Taguchi approaches G. Casa...

286KB Sizes 0 Downloads 44 Views

Journal of Materials Processing Technology 167 (2005) 422–428

Investigation on Ti6Al4V laser welding using statistical and Taguchi approaches G. Casalino a,∗ , F. Curcio b , F. Memola Capece Minutolo c a

Department of Mechanical and Operational Engineering, Politechnique of Bari, Viale Japigia,186, Bari 70126, Italy Department of Mechanical Engineering, University of Salerno, Via Ponte don Melillo, 1, Fisciano (Sa) 84084, Italy Department of Materials and Production Engineering, University of Napoli “Federico II”, P. Tecchio, 80, Napoli 80125, Italy b


Abstract Ti6Al4V is presently one of the most widely used titanium alloys, accounting for more than 50% of all titanium tonnage in the world, and to date no other titanium alloy has been a threat to its dominant position. Laser welding of Ti6Al4V is a major issue in the automotive and aerospace industries. In this paper, both CO2 and diode laser welding processes were investigated for Ti6Al4V alloy sheet joining using either lap or butt configurations. Artificial neural networks (ANN) processed the data coming from the experimental trials. The aim was to interpolate the database in order to form a suitable database for the analysis of the variance (ANOVA) and the Taguchi analysis of the means. The ANOVA has the merit of being able to be validated on a statistical basis while the Taguchi engineering approach has the ability of maximizing the information coming from a small database of data. Therefore, reliable and valuable information over the ranges of the processes investigated was individuated from a limited number of trials. © 2005 Elsevier B.V. All rights reserved. Keywords: Ti6Al4V; CO2 laser; Diode laser; Welding; ANN; Taguchi DOE

1. Introduction Ti6Al4V is an alpha–beta alloy that is the workhorse alloy of the titanium industry. The grade 5 alloy is the most commonly used alloy, over 70% of all alloy grades melted being a sub-grade of Ti6Al4V. Its uses span many aerospace airframe and engine component uses and also major non-aerospace applications in the marine, offshore and power generation industries in particular. The addition of 0.05% palladium (grade 24), 0.1% ruthenium (grade 29) and 0.05% palladium and 0.5% nickel (grade 25) significantly increase corrosion resistance in reducing acid, chloride and sour environments, raising the threshold temperature for attack to well over 200 ◦ C. The essential difference between Ti6Al4V ELI (grade 23) and Ti6Al4V (grade 5) is the reduction of oxygen content to 0.13% (maximum) in grade 23. This confers improved ductility and fracture toughness, with some reduction in strength. Grade 23 has been ∗

Corresponding author. Fax: +39 080 5962788. E-mail address: [email protected] (G. Casalino).

0924-0136/$ – see front matter © 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.jmatprotec.2005.05.031

widely used in fracture critical airframe structures and for offshore tubulars. Mechanical properties for fracture critical applications can be enhanced through processing and heat treatment. Grade 29 also having lowered level of oxygen will deliver similar levels of mechanical properties to grade 23 according to processing [1]. The success of titanium alloys in some manufacturing industries is due also to the rapid advance of titanium metallurgy and the successful solution of problems associated with the development of welding methods. These alloys may be joined soundly by a wide variety of conventional and solid-state processes although its chemical reactivity requires special precautions to avoid contamination of the fusion and heat-affected zones both on the face and root sides [2,3]. Fusion welding of titanium has been performed principally in inert gas-shielded arc and high energy beam welding processes. Among the high power density technologies, the electron beam and laser welding have showed a great capability in producing narrow and deep joints [4–6]. The amount of heat used in laser welding is roughly comparable to that of conventional arc welding processes. As the heat is focused

G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428

on a very small area the weld pool is much smaller than in arc welding. The welding speed is much higher, up to approximately the speed of conventional welding processes [7]. In practice, three types of laser are available for welding: CO2 , Nd:YAG and diode laser. Focusing of the laser light gives a high energy density making it suitable for welding operations. Depending on the energy density of the laser beam three characteristic types of weld pool modes can be discerned. At increasing energy density the weld pool changes from conduction mode to deep welding and then to keyhole mode. Both CO2 and Nd:YAG lasers operate in the infrared region of the electromagnetic radiation spectrum, invisible to the human eye. The Nd:YAG provides its primary light output in the near infrared, at a wavelength of 1.06 ␮m. This wavelength is absorbed quite well by conductive materials, with a typical reflectance of about 20–30% for most metals. The near infrared radiation permits the use of standard optics to achieve focused spot sizes as small as 25 ␮m in diameter. On the other hand, the far infrared (10.6 ␮m) output wavelength of the CO2 laser has an initial reflectance of about 80–90% for most metals and requires special optics to focus the beam to a minimum spot size of 75–100 ␮m in diameter. However, whereas Nd:YAG lasers might produce power outputs up to 500 W, CO2 systems can easily supply 10,000 W and greater. As a result of these broad differences, the two laser types are usually employed for different applications. The powerful CO2 lasers overcome the high reflectance by keyholing. The reflectivity of the metal is only important until the keyhole weld begins. Once the material’s surface at the point of focus approaches its melting point, the reflectivity drops within microseconds. Laser types can operate in either the continuous or the pulsed mode. CO2 lasers, which range in power from 50 to 15,000 W, are more efficient in their conversion of electrical power to laser radiation than Nd:YAG lasers, which range from about 50 to 800 W output power. However, as discussed above, the reflectivity of most metals is much higher at the CO2 wavelength than the Nd:YAG wavelength. Recent advances in fast-axial-flow CO2 lasers provide improved beam characteristics, making these systems competitive with electron beam welding for deep-penetration applications. Solid-state lasers (the generic name for Nd:YAG, Nd:glass and similar lasers) are preferred for low to moderate power applications. They have found extensive application in the electronic/electrical industries for spot welding and beam lead welding integrated circuits to thin film interconnecting circuits on a substrate [8]. High power diode laser welding makes use of the conduction heat transfer mode. The wavelength of the laser light is dependent on the concentration of dopes in the GaAs crystal, like aluminium (780–800 nm), indium (880–1100 nm) and phosphor (630–690 nm). Aluminum is commonly used for high power diode lasers. A typical wavelength is 808 nm. A PN semiconductor junction is capable of generating a laser power of several milliWatts. In order to be able to gener-


ate higher power levels a large number of PN-junctions are assembled creating a diode laser bar capable of generating approximately 50 W. To generate more power several tens of diode bars, fitted with micro-channel water cooled sinks are combined into a stack, capable of delivering approximately 1.5 kW. By combination of several stacks power levels of up to 6 kW can be generated. Due to its construction, the beam shape of the diode laser is not round and shows a large divergence. Further, the divergence is different in the parallel and perpendicular directions with respect to the bar. The brightness of the diode laser is lower compared to the CO2 and Nd:YAG laser. In order to mitigate this, per diode bar and per stack special optics like lenses, prisms and mirrors are used to deliver the optimum beam quality. As the laser light of the diode laser is originating from several resonators the resulting beam is not coherent. This is by no means a disadvantage in materials processing. Apart from direct delivery, fiber systems can also be used to deliver the beam of the high power diode laser, as in Nd:YAG lasers. Fiber diameters range from 0.4 to 1.5 mm. Advantages of the diode laser compared to CO2 and Nd:YAG lasers are the possibility to modulate the laser power (up to 10 GHz), high efficiency, low weight (10 kg), compact size (shoe box). The low unit weight and compact size permits integration of the high power diode laser without fiber into a robot system or a mobile material processing center. The relatively large focal spot and beam divergence make the diode laser a suitable candidate for material processing operations where extreme power densities are not required, as in soldering, conduction mode welding and surface treatment (transformation hardening and cladding). The efficiency of the laser source of 30% is very high and the stacks of the high power diode laser will last more than 10,000 h. This keeps the operational costs of the laser low. Furthermore, the capital costs of the high power diode laser are relatively low, roughly comparable to CO2 lasers. This keeps hourly rates low, allowing for a fast return on investment. A particular property of the (direct) diode laser is the rectangular spot of several mm2 cross-section. The lower power density that results gives a somewhat lower processing speed than conventional lasers. As a benefit the reduced demands on product tolerances and the resulting weld appearance more than compensates for this issue. Added to this are the reduced costs in part preparation and post weld operations, making the high power diode laser an attractive industrial laser [7]. This investigation aimed to test both the CO2 and diode laser welding of grade 5 Ti6Al4V alloy in the either lap and butt configurations.

2. Experimental data set interpolation using ANN As previous research has amply shown, neural networks are effective tools when it comes to modeling unknown or


G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428

sigmoid or hyperbolic tangent activation function: netl+1 = i


wij · xi

outl+1 = f (netl+1 i i )

Fig. 1. The architecture of a feed-forward type multi-layer neural network (MFNN).

semi-unknown processes [9]. In this case, artificial neural networks (ANN) were used in order to interpolate the incomplete experimental data set that was available for the analysis. The experimental data coming from the experimental trials are amended in order to be suitable for statistical analysis. As different network formats are proposed in the literature, the first problem to be solved is what type of network should be adopted. The results of previous analyses showed that the neural network used was the multi-layer feed forward type (MFNN) in 95% of the cases [10]. The frequent use of this type of network is mainly due to the fact that it acts as a universal interpolator in a wide range of cases in which input variables are to be related to a set of output variables. These networks are of the supervised learning type, in that the input/output relation can only be learned based on a set of training data (Fig. 1). The typical architecture of an MFNN is shown in Fig. 2. It includes an input layer (x), one or more hidden layers and an output layer. Each layer consists of one or more neurons linked together by weights used to convey data from all the neurons of the previous layer to all those present in the subsequent layer. The weight that links neuron i of layer l to neuron j of layer l + 1 is characterized as wij [11]. The input value for each neuron is determined as the weighted average of the output values of the neurons in the previous layer (l), whereas the output value is based on a




The implementation process of a neural network falls into two phases: the identification of the topology and the choice of the learning algorithm. Choosing a topology is a highly delicate process since the performance of the network will depend on it. In general terms, the greater the number of layers used and the greater the number of neurons in each of them, the more the network will approximate the test point cloud. However, as the neurons increase in number, the network’s generalization capacity is seen to decrease, so that the network will increasingly function as an associative memory X. To obtain an efficient network, the suitable quantity of layers and neurons will consequently have to be identified. The learning process [12] is conducted by showing the training set to the network on a continuous basis and adjusting weights until the error rate (3) decreases below a pre-defined value: ej =

1 d 2 (Y − Yj ) 2 j


The weights are adjusted based on the delta rule: wij (t) = −η

∂E + α · wij (t) ∂wij

wij (t + 1) = wij + wij (t)

(4) (5)

where η and α are the learning rate and momentum, respectively, with values between 0 and 1. However, the back-propagation algorithm based on the delta rule presents a shortcoming: convergence times are fairly long and the system risks getting stuck at local minimum values. An effective alternative is the Levenberg–Marquardt algorithm [13], which allows the learning process to be completed within a comparatively short timeframe, although the computing memory capacity needed is seen to increase in proportion to the complexity of the network.

3. CO2 laser welding 3.1. Experimental device and trials

Fig. 2. Network for AW.

A selection of laser beam butt and lap welds were performed using 2.85 mm thick Ti6Al4V annealed sheets and 2.5 mm thick sheets. A CO2 Rofin Sinar Mod. DC 025 (maximum power 2.5 kW) was used as a laser source. Although possible laser power ratings range between 0.5 and 2.5 kW, several power levels (from 1.8 to 2.4 kW) were tested at speeds ranging from 60 to 120 mm/s.

G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428

Laser power ratings and welding speed, as well as several different thickness values were analyzed as input factors, while morphological parameters (bead crown width, melting depth and melted area) and mechanical properties (mean micro-hardness values of the joint in the melted and heat-altered zones, as well as ultimate tensile strength) were considered as output variables. 3.2. ANOVA The data coming from the experimental trials needed to be interpolated in order to fill the blank levels in the ANOVA analysis of a three factors plan of the experiment. Therefore, a supervised back-propagation type neural network provided the complementary data. Input parameters were laser power rating, specimen thickness and welding speed. Output parameters were the average width of the melted area (AW), which was calculated as the ratio of the welded area to the actual thickness of the specimen, and the difference in micro-hardness between the base material and the welded joint (HV, FZ). The first output is related to the welding process efficiency and the second to the severity of the thermal cycle during welding. The plan that the authors used provided for 24 tests. The blanks corresponding to tests not performed were filled with the values provided by the neural network using specific input parameters. Figs. 2 and 3 show the architectures for the networks. In order to analyze these parameters, specimens were used that differed only slightly in thickness, which means that it was possible to adopt comparable laser power ratings and welding speeds. The following three factors were considered: Factor A: Welding speed (WS) – four levels – 60, 80, 100 and 120 mm/s. Factor B: Specimen thickness (T) – two levels – 2.5 and 2.85 mm, the latter was furnished in the annealed condition. Factor C: Laser power (P) – three levels – 1.8, 2.1 and 2.4 kW. The neural network assisted test plan was analyzed using a recent Minitab release [10]. Examining the results of the variance analysis [14] (see Table 1) it was found that the WS was significant at any confidence level, which is in line with the expectations. AW is not affected by thickness, i.e. the annealed state does not affect the size of the fused zone. AW

Fig. 3. Network for (HV, FZ).


Table 1 Results of analysis of variance Source

T P WS T×P T × WS P × WS


ANOVA per (HV, FZ)





3.23 3.65 134.7 2.23 8.58 1.55

0.123 0.092 0.000 0.189 0.014 0.305

13.13 5.49 3.78 6.01 2.75 1.71

0.011 0.064 0.078 0.057 0.135 0.265

increases, albeit slightly, in proportion to the power rating applied. These findings can be explained as follows: at the thickness values considered, a limited power range is sufficient to ensure in-depth penetration (1.8–2.4 kW) and within this range AW does not vary appreciably [15]. All the interactions but the WS one are statistically significant behind 99% level of confidence. As far as the two thickness values considered are concerned, i.e. 2.5 and 2.85 mm, it is evident that a different pre-welding condition affects the average micro-hardness value measured on the base metal. In particular, the average HV value measured on the 2.5 mm specimen is 335, while that measured on the 2.85 mm specimen is 290. This difference may also explain the significance of the thickness parameter below a 99% confidence level. From the analysis of the means, it can be said that as power is increased, the variation of micro-hardness decreases only at a 90% confidence level. If the remaining parameters are left unchanged, WS initially shows a declining trend and then a rising trend, with a (HV, FZ) minimum in the area where speeds approximate the critical value (the maximum speed at which in-depth penetration is observed). Globally speaking, the trends observed are in line with the results reported in the literature [6].

4. Diode laser welding 4.1. Experimental device and database description The laser used was the HPDL ROFIN SINAR DL 020 with a maximum power of 2.3 kW and focal lenses in ZnSe. Either argon or helium was used as the covering gas, spread by two nozzles. Both the crown and the root of the weld seam were protected from a distance varying between 2 and 5 mm. The nozzles were tilted at 25◦ angle with respect to the plan of the joint. The laser beam was focused into a rectangular spot of dimensions 1.8 mm per 6.8 mm on the upper part of the junction. Such a spot was positioned so that the 6.8 mm side was parallel to the principal direction of the joint. In order to find an optimal combination of the variables the experimentation must be robust which means insensitivity of control factors to the noise factors. Table 2 contains the control factors. The welding speeds values are shown in


G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428 Table 5 Levels of defects for each experiment

Table 2 Control factors Cause

Low level (1)

High level (2)

Gas Thickness Welding speeda

Helium (1) 1 mm (1) Slow (1)

Argon (2) 1.5 mm (2) Fast (2)


See Table 3.

Table 3 Welding speeds Welding speed (mm/min)

1 mm helium

1.5 mm argon

Slow Fast

500 800

400 950

Table 3. The responses of the experiments were the excess weld metal, the excessive penetration and the summation of the projected area of the porosity in a direction parallel to the surface and perpendicular to the weld axis. The defects were measured according to the EN ISO 13919-1 and -2 suggestions for laser welding of steels and aluminum alloys levels of imperfection, respectively [15,16]. The first two defects are related to an extra heat input and their profiles can generate notch effects in the crown and the root of the seam. The third defect is related to the welding conditions and weld pool shielding against gases. Lack of mechanical strength is due to a large porosity. The welded specimens were chemically etched using Keller’s reagent (2 ml HF, 10 ml HNO3 and 88 ml deionized H2 O). The defects were measured using a digital camera mounted on an optical microscope with a local reference coordinates system for scaling. 4.2. Taguchi data analysis A L4 plan tested the gas, the thickness and the welding speed effects on the levels of imperfections. Software for statistical analysis performed the calculations [17]. The different combinations of control factors levels are given in Table 4. Only the internal array was considered and each experiment was repeated twice in order to evaluate the noise effect. All the defects must be as small as possible in order to have a sound weld. Therefore in the experiments the quality characteristics are continuous and non-negative, which is they can take any value from 0 to ∞. The most desired value is zero. Such problems are characterized by the absence of a scaling factor or any other adjustment factor in the defect function.

1 2 3 4 1 bis 2 bis 3 bis 4 bis

Excessive metal (mm)

Excessive penetration (mm)

Porosity (mm2 )

0.08 0.08 0.05 0.19 0.10 0.10 0.12 0.06

0.06 0.07 0.04 0.06 0.00 0.04 0.00 0.07

0.0055 0.0085 0.0049 0.0400 0.0527 0.0057 0.0667 0.1318

Therefore, this is a typical static “smaller-the-better” problem. In this kind of problem, the S/N ratio is determined by the following expression:  n   S (6) = −10 log10 yi2 N i=1

where yi is the ith observation in Table 5. The mean for each level is calculated as the average of all the responses that were obtained for that level. The-smaller-the-better criterion should be used when the goal of the experimenter is to minimize the response of the plan [18]. Expression (6) gives the weld quality loss expressed in decibel in terms of defects and must be maximized in order to reduce the effect of noise. The analysis of means for each of the experiments will give the better combination of parameters levels that ensures a low level of defectiveness in the weld according to the experimental set of data. The plots in Figs. 4–6 account for the factor effects on the means of the measured levels of the defects. Each figure is portioned into three parts. Each part is related to a control factor, which is indicated in the upper side of the part, which is from left to right the gas type, the joint thickness and the welding speed, respectively. The lower side shows the levels for each control factor (1 or 2, see Table 2 for their meaning) whose mean is scaled on the vertical axis. The levels that produced the lower levels of the defects are encircled and then collected in Table 6. Table 6 resumes the best combination of control factors levels that respect the weld soundness criterion of eval-

Table 4 Plan of the experiments Conditions

Levels Gas



First experiment Second experiment Third experiment Fourth experiment

1 1 2 2

1 2 1 2

1 2 2 1

Fig. 4. Means responses for excess weld metal.

G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428


Fig. 7. S/N for excess weld metal.

Fig. 5. Means responses for excessive penetration.

Fig. 8. S/N for excessive penetration. Fig. 6. Means responses for porosity projection.

uation, which is the-minimum-the-best. The levels of the variables are indicated in the normalized forms either 1 or 2 (see Table 1 for the corresponding actual values of the parameters). Therefore, it was found that helium gives a lower level of the imperfections, at least for those considered in the investigations, then argon. The 1 mm thick weld showed smaller defects. Low speed is preferable. The results indicate that the three kinds of defects follow the same trend, which permits the affirmation that the three target function are synergic. The S/N ratios measured the effect of noise, i.e. the robustness of the response to the variability of noise effects. The general nature of the plots is the same as that for the means effect. Hereby, the encircled levels of control factors indicate less sensitiveness of the levels of defects to the noise factors (Figs. 7–9). Almost all the levels of parameters in Table 6 are also those who maximize the insensitiveness to the noise factors. The only warning is given by the effect of the thickness on the porosity projection. In this case, a lower sensitiveness would be obtained with the smaller thickness.

Table 6 Best setting for control factors (means) Type of defect



Welding speed

Excessive metal Excessive penetration Porosity

1 1–2 1

1 1 1

2 2 2

Fig. 9. S/N for porosity projection.

5. Conclusions In the first part of the paper, a database on butt and overlap welds from in-field and calculation trials was used to perform a statistical analysis of the effects of welding parameters on the shape of the welded area (AW) and the hardness variation of the base metal versus the welded area (HV, FZ) in Ti6Al4V alloy. Thanks to the interpolation of the experimental data by means of ANN, important information confirming the statistical significance of some parameters and interactions was provided by both the ANOVA and Taguchi approaches. The results of the ANOVA analysis matched those available in the literature for CO2 laser welding. In the second part of the paper, the application of the Taguchi L4 experiment to the HPDL welding of the titanium alloy Ti6Al4V sheets enabled the finding of the combination of some of the major welding control factors such as shielding gas and welding speed that produce low level of defects. It has been found that the levels of these imperfections are


G. Casalino et al. / Journal of Materials Processing Technology 167 (2005) 422–428

lower for the smaller thickness, the higher welding speed, and when the helium is used as shielding gas. The S/N ratio showed that these levels of the control factors are robust to the noise factors. Finally, the outcomes of this investigation validated either the ANOVA and Taguchi numerical approaches that were able to upgrade the knowledge of the two welding processes without the need of further experimental trials. In fact, results from ANOVA are more reliable thanks to the solid statistical basis on which the evaluation models are built. On the other hand, using the Taguchi philosophy of data analysis, the evaluation is rapid and performed intuitively and the optimization process is performed empirically.




[7] [8] [9]

Acknowledgments The Welding Department of the Silesian University of Technology in Gliwice, Poland, is acknowledged for operating the diode welds. Senior researcher Giuseppe Daurelio is also greatly acknowledged for his collaboration during the CO2 welding trials.




[1] Titanium: The Ultimate Choice, Titanium Information Group, [2] L.S. Smith, M. Gittos, P. Thereadgill, High quality and productivity joining processes and procedures for titanium risers and flowlines, in: Sintef (Ed.), Titanium Risers and Flowlines Seminar, Norway, 1999. [3] F. Curcio, F. Memola Capece Minutolo, A. Klimpel, A. Lisiecki, A basic study on the high power diode laser welding of a titanium alloy, in: Proceedings of the 23rd ICALEO, International Congress

[11] [12] [13]

[15] [16]

[17] [18]

on Applications of Lasers & Electro-Optics, San Francisco, USA, 2004, pp. 25–32. Q. Yunlian, D. Hu, H. Quan, Z. Liying, Electron beam welding, laser beam welding, and gas tungsten arc welding of titanium sheet, Mater. Sci. Eng. A280 (2000) 177–181. F. Caiazzo, F. Curcio, G. Daurelio, F. Memola Capece Minutolo, Ti6Al4V sheets lap and butt joints carried out by CO2 laser: mechanical and morphological characterization, J. Mater. Process. Technol. 149 (2004) 542–548. J.L. Barreda, F. Santamaria, X. Azpiroz, A.M. Irisarri, J.M. Varona, Electron beam welded high thickness Ti6Al4V plates using filler metal of similar and different composition to the base plate, Vacuum 62 (2001) 143–150. C.B. Miller, Laser Welding Article, Laser Corporation, W.N. Steen, Laser Materials Processing, Springer-Verlag Ltd., London, 2003. B.K. Wong, T.A. Bodnovich, Y. Selvi, Neural network applications in business: a review and analysis of the literature (1988–1995), Decis. Support Syst. 19 (1997) 301–320. Minitab User’s Guide 2: Data Analysis and Quality Tools, Minitab Inc., 2000. T. Khanna, Fundamentals of Neural Networks, Addison-Wesley, 1990. S. Haykin, Neural Networks: A Comprehensive Foundation, second ed., Prentice Hall, 1998. ¨ B.M. Wilamowski, O. Kaynak, S. Iplikci, M. Onder Efe, An Algorithm for Fast Convergence in Training Neural Networks, IEEE, 2001, pp. 1778–1782. D.D. Montgomery, Design and Analysis of Experiments, fifth ed., John Wiley & Sons, New York, 2001. EN ISO 13919-1, Electrons and Laser Beam Welded Joints, Guidance on the Levels for Imperfections, Part 1: Steel, 1996. EN ISO 13919-2, Electrons and Laser Beam Welded Joints, Guidance on the Levels for Imperfections, Part 2: Aluminum and its Weldable Alloys, 2001. MINITABTM Statistical Software Release 13—Perpetual License, USA, 1989. M.S. Phadke, Quality Engineering using Robust Design, PTR Prentice Hall, Englewood Cliffs, NJ, USA, 1989.