Misalignment detection for friction stir welding or

Enabling seam tracking for friction stir welding

Paul A. FlemingDavid H. Lammlein

D. M. WilkesGeorge E. CookAlvin M. StraussDavid R. DeLapp

Daniel A. Hartman

April 2008

Abstract

This paper describes a technique for determining the position of thefriction stir welding (FSW) tool with respect to the weld seam duringwelding. This technique is proposed for online misalignment detection oras a position estimator for in-process tracking of the weld seam for FSW.

1 Introduction

1.1 Friction stir welding

Friction stir welding (FSW) is a method of welding where material is joined bya rotating tool which traverses along the joint line [1]. It was first patentedin 1991 by The Welding Institute and has since found an increasing number ofapplications [2]. The basic setup of FSW is shown in Fig. 1.

Fig. 1 illustrates an FSW butt-joint, where the material to be joined arebutted together end to end. The FSW tool, consisting of a pin (or probe) andshoulder rotates and traverses the joint, applying heat and stirring the materialtogether [1]. Notice that the two sides of the weld are named based on whetherthe side of the tool is rotating with the welding direction (advancing side) oragainst (retreating side). This nomenclature is used throughout this paper.

Butt-welds however are only one type joint to which FSW is applied. Anumber of other joint types have been shown to be applicable for FSW, includ-ing: single lap welds and multi-lap welds, 2 and 3 piece T-joints, edge butts andcorner fillet welds [3]. Fig. 2 illustrates two of these other joints used in FSW:lap weld and T-joint.

1

Figure 1: The essential schematic diagram of FSW from [1]

Figure 2: Lap FSW (left) and T-joint FSW (right)

2

In these types of FSW, the probe penetrates through the upper member andinto the lower member, stirring the material together to form the joint.

1.2 FSW alignment and seam tracking

This paper discusses techniques for misalignment detection for FSW. This tech-nology could be useful as a means of in-process monitoring to ensure the tool isproperly aligned throughout the weld. Additionally, this could be incorporatedas feedback to implement seam tracking for FSW. Typically in the FSW liter-ature, the weld seam is linear and the FSW process is rigidly clamped so thatthe alignment of the tool to the seam can be set at the start of the weld andwill remain aligned throughout the weld. However, there are also technologiescapable of implementing non-linear weld-seams. Robotic FSW is one such tech-nology. The apparatus described in [4] is another, which allows a sheet of metalto be guided through FSW either by an operator, or else by controlled actua-tors. These technologies do not intrinsically follow the weld seam, and rely onoperator control, either by hand or by joystick, preprogrammed welding pathsor computer vision to track the non-linear weld seam. The technology describedin [5], which uses a tactile sensor at the end of a robotic manipulator to traceand record a 3-dimensional weld seam, could be used to learn the weld seamahead of time and be used as the programmed input to the above technologies.

In all FSW joint types, the alignment of the FSW tool with respect tothe weld seam is important to ensure good weld quality. In butt welds, animproperly aligned tool can result in root flaws [6]. In the extreme, a severemis-alignment will result in no weld at all if the tool is entirely located in onlyone sample. However, the effect of misalignment and its severity is dependenton weld configuration and other parameters.

T-joints are particularly susceptible to misalignment because the weld lineis not observable from above. In T-joints, the effects of offset depend upon thedimensions of the material used, the dimensions of the FSW tool employed andthe clamping methods used. In one method of clamping for T-joints, steel blocksare clamped alongside the vertical member and under the horizontal member.These blocks provide rigidity for the vertical member. If the corners of the blocksnear the weld are sharp, then they allow little material to escape. However, ithas been shown that machining a radius into this corner and allowing materialto be ejected into a small fillet improves weld quality [7].

One case in which the weld is particularly sensitive to offset is the “open-air” clamp, where there is effectively space alongside the contact plane of thehorizontal and vertical member. In this case, if the probe is offset from thecenter of the vertical member, material is ejected into this space leaving voidsin the weld.

In this work, T-joint FSW with “open-air” clamping is used as a test-bedto demonstrate misalignment detection. This fixturing setup is selected bothbecause of its sensitivity to offset and its practical implications in industry.However, it is expected that other users may prefer clamps with small fillets, aswas done in [7] and [8]. The approach outlined in this paper is equally applicable

3

to these clamping schemes as well.

2 Force as a process feedback mechanism

In this work, force values are used to determine offset position. Using force as afeedback signal is typical in the literature of monitoring and control of frictionstir welding. For example, the paper “Neural network evaluation of weld qualityusing FSW feedback data” uses the frequency content of the dynamic forcesignal to discover metallurgical defects and evaluating tensile strength duringthe welding process [9]. The paper “Analysis of the FSW force footprint andits relationship with process parameters to optimise weld performance and tooldesign” demonstrates a relationship between forces experienced by the tool andprocess parameters such as welding and rotation speeds, as well as tool geometry[10].

Additionally, axial forces are typically monitored and controlled to ensureweld quality (load control) [11][12]. It has been shown that a minimum axialforce is essential to ensure generation of sufficient frictional heating. The axialforce is monitored by force sensors, and maintained by either changes to thevertical position of the tool, rotation speed or welding speed.

In this paper, it is demonstrated that force values can be used to monitorand control alignment. However, in this case of alignment detection, there isnot a simple relationship between weld forces and tool alignment. Neverthelessit is possible to develop estimators which can predict offset position given forcedata.

For force collection, a Kistler dynamometer is used. This apparatus can readaxial (z) force, planar forces (x and y) as well as the moment about the z-axis.Optical interrupters were employed in order to determine the rotational angleof the tool and dynamometer relative to the welding sample at a given reading.Using this setup, one “force sample” is composed of 40 force readings: each ofthe 4 force readings taken at 10 different angular positions through one rotation.This is illustrated in Fig. 3

Figure 3: The fundamentals of a force sample using a Kistler dynamometerduring FSW

4

3 Experimental setup

3.1 Experiment

In order to develop an offset estimator for T-joints, 30 T-joint welds were runwith offsets ranging from 4mm offset to either side in incremental steps of0.25mm.

Figure 4: Schematic diagram of T-joint welding configuration used in this paper

The setup of the T-joints is illustrated in Fig. 4. Both the horizontal andvertical members are 6061 aluminum, with the horizontal member measuring3.175mm in thickness and the vertical member being 9.525mm across. Theclamps were steel, with a 3mm x 3mm notch milled in the top to simulate“open-air”. Though not shown, the horizontal member was also clamped down.Finally, the FSW tool consisted of a 5mm diameter 3.81mm long threaded probeand 19mm diameter shoulder. The rotation speed was fixed at 1000 RPM, andthe weld speed at 100mm min−1.

Force samples were recorded according to the method described earlier.These forces were then inspected to determine if any forces exhibited a cor-relating relationship with the changing offset. It was discovered that indeedsome forces did demonstrate this relationship. In some cases, this relationshipwas simple. For example, the axial force generally increased as the offset magni-tude approached 0. This relationship existed regardless of the angular positionof the dynamometer relative to the material.

In Fig. 5 the recorded axial forces for each run are plotted against the offsetof the tool for the weld in which they were recorded. The forces are organized

5

into a boxplot, where each box’s edges represent the upper and lower quartile ofthe data, the middle line the median, the lines, or whiskers, extend out to smallor forces which are not outliers, and finally the crosses indicate outliers [13].This plot style is used to indicate the distribution of axial forces throughouteach weld, with approximately 1000 force readings used from each weld. Alsoshown are cross-sections of some of the welds corresponding to the same offsetvalues. These are included to demonstrate representative cross-sections of thewelds at given offsets. The cross-sections differ largely through the presence orabsence of voids and material expulsion.

Figure 5: Comparing axial forces and offset

A similar relationship is observed in torque. However, when looking at therecorded x and y (planar) forces, there is a more complicated relationship. Therelationship is dependent on the angular position of the dynamometer at theforce recording. Some angles produce very useful relationships while other showlittle correlation. Fig. 6 demonstrates the relationship of x-force values recordedat the first rotational position vs. offset as was done in Fig. 5

A useful property of the planar forces is that, (as seen in Fig. 6), the forcesallow for differentiation of a tool offset to the advancing side versus one offsetto the same degree to the retreating sides. Combining the information fromthe axial force with that of the planar force allows for the determination ofabsolute position of offset (versus merely detecting magnitude of offset withoutdirection). Because of this, it is possible to develop a complete position whichpredicts both offset direction and magnitude accurately.

3.2 Development of offset position estimator for T-joints

Based on these signals, an estimator was developed to predict offset given theseforce samples. A general regression neural network (GRNN) was selected to

6

Figure 6: Comparing planar forces and offset

accomplish this. A GRNN is an artificial neural network which estimates con-tinuous variables using non-parametric estimators of probability density func-tions [14]. A principle advantage of using a GRNN is that it converges to theconditional mean regression surface and can form “very reasonable” regressionsurfaces with only a few samples [14].

In this experiment, thirty welds were performed, each with a different offset.Each weld run generated approximately 1000 force samples (40 dimensionalvectors of collected forces). Using “backward wrapping”, the dimensionality ofthe samples was reduced from 40 to 15. In backward wrapping, features areremoved one at a time, and their effect on the performance of the classifieris observed. Features are removed until the classifier performance ceases toimprove.ed. Features are removed until the classifier performance ceases toimprove.

Next the network was repeatedly trained and tested using “leave-one-outcross-validation”. Cross-validation estimates how well the network will performon unseen data [15]. In this method, one run was removed from the data, thenetwork was then trained using the remainder of runs. The trained network wasthen used to predict the offset of the held-out run for each force sample. Theresults were recorded, and then the process repeated for each weld run. Theresults of this experiment are shown in Fig.7.

Shown in Fig.7 is the prediction of the samples of each run organized intoa boxplot. Each box is composed of approximately 1000 position predictions.The estimator averaged an absolute error of 0.42mm.

7

Figure 7: Offset predictions of GRNN

4 Results

To demonstrate the capabilities of using this technique for tracking a weld seam,a number of welds were run where the offset position varied during welding. Theactual values of the offset position were recorded. Next the collected force datawas applied to the estimator network and its predicted offset for each forcesample were compared with the true position. Some of these results are shownin Fig.8 and Fig.9. The figures show the true offset position of the FSW toolwith respect to the weld-seam, along with the predicted. Also illustrated forcomparison is a green region which marks the region of offset values which didnot contain a void.

In Figure 8, the probe starts offset to the advancing side of the weld, itthen shifts into the void-free region of offset positions, and finally shifts intothe retreating side. The offset position estimates are close to the true positions.However, the lateral motion causes force disturbances which affect the predictioncausing the overshoot around 25 seconds as well as the drop around 60 seconds.This effect will need to be considered in any real-time control algorithm basedon this method of offset prediction.

In Figure 9, the tool advancing side of the weld free region, to the retreatingside in order to demonstrate the ability of the estimator to track within oneregion type. It then shifts quickly out of the weld free region to the advancingside. Although the estimator does predict the weld is too far off to the advancingside, there is a degree of error. This indicates that thirty welds is probably

8

Figure 8: Predicted and actual offsets over time

9

too few to embody the relationship between offset and forces and likely moreexamples are needed to further refine the estimator.

Figure 9: Predicted and actual offsets over time

5 Discussion

5.1 Discussion of results

As demonstrated in the results, a reasonably accurate estimator was built forT-joint FSW. The accuracy of the estimator could very likely be improved signif-icantly by inclusion of a larger sample of runs at each offset position. The resultspresented clearly show that feedback control can be implemented to maintainproper positioning over the joint well within the window of acceptable welds.

The techniques employed in this paper for determining offset position inT-joint FSW can be applied to the other joint types. The method functionsby discovering the way in which offset affects weld forces and regressing to afunction. In the case of T-joints, changing offset position affects a number ofphysical characteristics which in turn affects forces. For one example, withincreasing offset from the center, the composition of material directly below theshoulder changes from mainly aluminum to including the steel clamps and air

10

gaps, which likely affect the axial force. Additionally, the proximity of the probeto the edge of the vertical member increases the likeliness of material expulsionwhich duly affects force.

Although anecdotal, these physics-based manifestations exist in the otherFSW joint types. In the event that changing offset positions produces onlymild changes in forces, there is always the possibility of adding features such asgrooves or elevations to the material or backing plate to augment the signals.Finally, in the event that offset produces equivalent changes when offset in eitherdirection, then a weaving method could be used to gain the center position.

6 Conclusions and future work

This continued refinement and adaptation of this research can be of great benefitto the deployment of FSW in more applications. Future work will focus on thedevelopment of a complete closed-loop seam tracking system, which will bebased on the research disclosed in this article.

This research is currently the subject of a US patent application.

References

[1] R. S. Mishra and Z. Y. Ma, Materials Science & Engineering R-Reports,2005, 50(1-2), Iii–78.

[2] W. M. Thomas, E. D. Nicholas, J. C. Needham, M. G. Murch, P. Temple-Smith and C. J. Dawes: G.b. patent application no. 9125978.8, 1991.

[3] C. J. Dawes and W. M. Thomas, TWI Bulletin, 1995, 120–128.

[4] T. W. Nelson, C. D. Sorensen, P. W. Carter, T. L. K. Jr. and D. V.Kirkham: Apparatus and method for performing non-linear friction stirwelds on either planar or non-planar surfaces, us 10161295, 2007.

[5] R. J. Heideman, E. G. Blachowiak, C. B. Smith, S. L. Carian, B. S. Tarr,N. A. Duffie and M. H. Predith: Three dimensional tactile seam tracingdevice, us 6013997.

[6] A. J. Leonard and S. A. Lockyer: ‘4th International Symposium on FrictionStir Welding’, Park City, Utah, USA, defects and Faults.

[7] K. Erbsloh, C. D. Donne and D. Lohwasser, Materials Science Forum, 2003,426-432, 2965–2970.

[8] L. Fratini, G. Buffa, L. Filice and F. Gagliardi, Proceedings of the I MECHE Part B Journal of Engineering Manufacture,, 2006, 220(5), 669–676.

[9] E. Boldsaikhan, E. Corwin, A. Logar and W. Arbegast: ‘Friction Stir Weld-ing, 6th International Symposium’, Saint-Sauver, Nr Montreal, Canada.

11

[10] D. G. Hattingh, T. I. v. Niekerk, C. Blignault, G. Kruger and M. N. James,Welding in the world, 2004, 48(1/2), 50–58, sensing.

[11] J. D. Kinton and J. Tlusty: Method and apparatus for controlling down-force during friction stir welding, us 6050475, 2000.

[12] T. V. Stotler and T. J. Trapp: Friction stir welding travel axis load controlmethod and apparatus, us 7,216,793 b2, 2007.

[13] J. Devore and N. Farnum: ‘Applied Statistics for Engineers and Scientists’,Duxbury Press, Pacific Grove, CA, 2nd edition, 1999.

[14] D. F. Specht, IEEE transactions on neural networks, 1991, 2(6), 568–576.

[15] S. J. Russell and P. Norvig: ‘Artificial Intelligence: A Modern Approach’,Prentice Hall, Upper Saddle River, NJ, 2nd edition, 2003.

12