TECHNICAL ARTICLE

Piping and Pressure Vessel Welding Automation throughAdaptive Planning and Control

SAM ROBERTSON ,1 JOSH PENNEY,1 J. LOGAN MCNEIL,1

WILLIAM R. HAMEL,1,3 DAVID GANDY,2 GREG FREDERICK,2

and JON TATMAN2

1.—Mechanical, Aerospace, and Biomedical Engineering Department, Tickle College of Engineering,University of Tennessee, Knoxville, Knoxville, TN 37996, USA. 2.—Welding Technology and RepairCenter, Electric Power Research Center, Charlotte, NC 28262, USA. 3.—e-mail: whamel@utk.edu

Welding automation is a pathway to reducing costs in the energy sector anddependence on certified welders, who are in short supply. Recent research intosystem-level automation of multibead/layer Tungsten Inert Gas welding forstainless-steel components is presented. Automation is pursued for weldplanning, execution, and defect detection. Planning utilizes active seam/-groove sensing and intuitive weld bead positioning. Bead and layer geometriesare estimated using weld bead shape prediction that takes into account pro-cess parameters. After the first weld layer, subsequent trajectory plans areadapted to compensate for differences between the planned and actual weldsurface. Sensor-based, closed-loop control of the process is being pursued tocompensate for gravitational effects. Continuous monitoring of the real-timeprocess to predict/detect the occurrence of welding defects is in development.Near-real-time defect detection provides the opportunity for immediate eval-uation and correction, reducing costly repairs. Preliminary experimental re-sults are presented.

INTRODUCTION

Industrial welding is an area of manufacturingthat has been highly automated over recent dec-ades. Operations previously performed by humanwelders are now completely handled by industrialrobots in many modern applications. Doing soremoves humans from more hazardous workingenvironments and frees up workers for tasks thatare not suitable for robots.

In current manufacturing of pipe and reactorpressure vessel (RPV) structures, mechanizedapproaches utilizing automated torch positioning areemployed (orbital typically) but rely primarily onmanual and supervisory control by skilled welders.Such reliance on human welders poses a number ofissues for the overall process. First, production capac-ity is limited by the availability of skilled welders, aseach welding system requires a dedicated humanoperator. This need is exacerbated by the current andprojected future shortage of qualified and certifiedwelders.1,2 Additionally, human operators are not well

suited for tedious and continuous process monitoringfor quality control judgments, particularly those asso-ciated with potential subsurface weld defects. Thisrequires postprocess nondestructive evaluation(NDE) to locate subsurface defects which, if found,require costly and lengthy repair measures before thewelded structure can be certified.

Therefore, a higher level of welding processautomation—one that maximizes the efficiency ofhuman operators and reduces the necessity forpostprocess repair—would significantly benefit pip-ing and pressure vessel manufacturing in all typesof power plants. To achieve this level of desiredautomation (worker optimization and repair reduc-tion), two categories of automated functions must bedeveloped: one that reduces control functions cur-rently performed by humans, and one that seeksnear-real-time weld evaluation and defect predic-tion/identification/correction.

The next section reviews the considerable amountof prior research in the areas of sensing and controlthat are relevant to the research aims of this work.

JOM, Vol. 72, No. 1, 2020

https://doi.org/10.1007/s11837-019-03912-y� 2019 The Minerals, Metals & Materials Society

526 (Published online November 22, 2019)

Many elements of this prior work provide importantfoundations and insights, but none have been foundthat address the overall system-level automationthat is sought in this work.

PRIOR RESEARCH

Significant research has been conducted onrelated subjects such as weld modeling, sensing,and control. An overview of weld modeling litera-ture compiled by Babu3 provides a thoughtfulcollection of research on modeling. A noticeableissue in current modeling and simulation researchis the computational throughput requirements,with the fastest published models running atslightly subminute total execution time scales forsimple bead-on-plate simulations.4 All of theapproaches to solving the rigorous governing equa-tions result in computational requirements thatcannot be met with the types of embedded comput-ing associated with modern digital control systems.

Similarly, research in the areas of sensing andcontrol for Tungsten Inert Gas (TIG) welding ismainly limited to simple, single pass types of welding.Additionally, research is limited to the control of oneaspect of welding (oftentimes penetration)5,6 insteadof treating the holistic process. Literature that aimsat more general process control often provides little orambiguous results.7 A major focus of sensor- andmodel-based control research has been vision based,yet it is intuitive that thermal sensing should proveprofitable, while some researchers as well as skilledwelders assert the usefulness of acoustic sensing.5

Hence, research into multisensor-based control of theholistic welding process for complex, multipass,multilayer operations is lacking but is needed fortechnological progression.

An additional consideration for full weldingautomation is the complexity added by orbitalwelding. As welding is conducted circumferentiallyfor pipes and pressure vessels, or in any operationthat requires welds to be performed out of align-ment with gravity, gravitational effects on the fluidflow further complicate modeling, sensing, andcontrol of the welding process as the processdynamics change. These changes are drastic enoughthat the ASME Boiler and Pressure Vessel Code(ASME BPVC) contains different qualification stan-dards for welds conducted in a variety of non-gravity-aligned (NGA) or ‘‘out of position’’ welds.8

Interest in orbital welding has been consistent fordecades due to the production benefits;9,10 however,discussion of control in orbital welding scenarios isgenerally lacking in literature. Even in researchspecifically aimed at good welding control results,discussion on gravity compensation in NGA weldingpositions is omitted.11

The remaining sections of this paper are orga-nized as follows: ‘‘Welding Automation Concept’’section focuses on the functional elements of sys-tem-level welding automation, ‘‘Experimental

System Description’’ and ‘‘Current Results’’ sectionscover the experimental setup and results obtainedthus far, and ‘‘Summary and Future Work’’ sectionprovides a summary and discussion about futurework.

WELDING AUTOMATION CONCEPT

For comprehensive system-level welding automa-tion, a sensor- and model-based control system thatcan compensate for dynamic events and distur-bances in the welding environment is needed. Thegeneral functional architecture of such a system isshown in Fig. 1. Three integrated functionalstreams are depicted: multibead/layer weld plan-ning, adaptive weld plan execution and control, andparallel weld defect detection. Parallel near-real-time weld defect detection seeks to identify possibledefects as they occur, giving operators the opportu-nity to determine further actions quickly.

Weld Planning

Weld planning is the process of determining thebaseline set of beads and layers necessary for aparticular job and generating the correspondingweld head/robot trajectories that will place thebeads in the proper sequence and locations withrespect to the workpiece. To begin, the geometry ofthe groove or seam to be welded is digitized using aprofilometer (a Keyence blue laser profilometer inthis case). Since the target of this work is new

Fig. 1. System functional diagram.

Piping and Pressure Vessel Welding Automation through Adaptive Planning and Control 527

construction fabrication, it can be safely assumedthat the prewelded groove geometry is fairly con-sistent, i.e., that the cross-sectional geometry at anypoint or segment is representative of the cross-sectional geometry at any other point or segment.This mitigates the need to image the entire joint forplanning purposes or to create a complex three-dimensional (3D) plan, utilizing instead a simplethree-dimensional plan that simply extends thetwo-dimensional (2D) section profile along thelength of the welds (Fig. 2).

At this point, no welds, other than tack welds,have been performed. Naturally, groove distortionswill occur during welding due to thermal expansion,and fit-up deviations may be present. However,these will be handled by the adaptive weld execu-tion control when the results of a specific weld layerare compared with the baseline plan and necessarycorrections to the next layer’s plan are made.

Control-Oriented Weld Bead Modeling

With the groove geometry known, the volume tobe filled can be determined as well as the volumeand geometry of the cold base metal. This informa-tion is combined with the weld process parametersto generate estimations of the weld shapes andcross-sections based on their placement inside thegroove (Cartesian position and torch angle in the

workpiece space). To produce weld shape estima-tions within the computational framework of thecontrol system (desktop computer processing rates),a relatively simple weld bead mathematical model isneeded. While numerical solutions of the detailedmultiphysics governing equations have been used tosimulate welding processes in the past, such modelscan take minutes to hours on high-performancecomputers to simulate a single bead-on-plate weld.4

For weld planning in this context, on the order of 30welds must be simulated in seconds. Therefore, a‘‘control-oriented’’ parametric bead model that canbe realistically executed within the control systemembedded computing environment has been devel-oped (and continues to be refined) to link weldparameters and torch tip orientation—with respectto the workpiece—to resulting weld shapes with anaccuracy that is sufficient for initial planning andlayer-to-layer adaptive corrections.

The automatic job planning algorithm uses thecontrol-oriented weld bead geometry model and theempty groove geometry to specify the position andorientation for each weld bead within its layer asneeded to fill the groove to completion. Weld vol-ume, groove volume, tie in, penetration, and overlaprequirements are used as inputs to drive an itera-tive algorithm that determines the number of weldlayers required to fill the groove and the optimal

Fig. 2. Weld planning tools and visualization: (a) Keyence profiles stitched into a 3D profile for an empty groove, (b) and a partially weldedgroove; (c) Graphical weld planning tool; (d) Initial parametric weld model.

Robertson, Penney, McNeil, Hamel, Gandy, Frederick, and Tatman528

bead placement for each weld layer. The results arereturned to the system operator for review/modifi-cation and approval. Upon approval, the planningsystem automatically generates the correspondingtorch trajectories that are ultimately used as robotservo commands to the Liburdi weld head.

Figure 2 shows the current state of the weldplanning and execution interface. For initial work, asimple ellipsoidal weld bead model utilizing beadheight and width has been used. However, aparametric model based on the welding parametersand an assumption that abutting parabolas form areasonable bead profile/penetration estimate hasbeen developed. This model is currently beingexperimentally evaluated, but utilizes the inputpower and material flow characteristics of the weldprocess to generate the upper and lower parabolasthat estimate the cross-sectional shape of the weldbead. Work by Tatman et al.12,13 is the basis for thismodel; initial results are promising.

Adaptive Weld Plan Execution

Welding operations will begin with the torch tiptrajectories from the initial plan. Since the planningprocess is based on approximate weld bead models,the actual final layer surface will most likely bedifferent from the surface represented in the plan.The plan for the next layer will be adapted tocompensate for the previous actual/planned differ-ence based on profilometer characterization of theactual layer geometry. In this way, near-real-timefeedback of the actual previous weld layer surfacegeometry is used to evaluate how well the groove isbeing filled and the variance between the ‘‘planned’’and actual welds. The adaptive bead path algorithm(in development) uses inputs from the torch headsensor suite to monitor how the groove is being filled(Fig. 3). The profilometer, mounted a few inchesahead of the torch, digitizes the upcoming groove

geometry, which is used to adjust the torch trajec-tory plan, ensuring that welds maintain the correctrelative position with respect to the groove featuresand other welds. In addition, knowledge of the torchposition combined with the profile scans of priorwelds can be used to identify actual (as opposed toestimated) weld geometries. Significant differenceswill be used to update the bead geometry model andadjust the idealized plan accordingly, resulting incurrent and future trajectory changes for the weld-ing operations. In extreme scenarios, welds may beadded or removed from the plan, but in any scenarioin which a significant plan modification is made, theoperator is notified by the planning window of theoperator interface. When significant differencesoccur related to the weld bead geometry, the actualgeometry will be fed into an algorithm to update theweld bead model.

For both planning and adaptation purposes, aslayers are estimated and evaluated, the smoothnessand average height of welded layers will be used tomake bead path plan adjustments. The averagesurface height of a given layer should, based ontypical weld layer geometries within grooves, beconsistent and rather flat. Therefore, for planningpurposes, the automated planning routine utilizesthe average weld height of the previous layer as theflat surface on which to plan the subsequent layer.During in situ analysis, the actual average heightand deviation from the bead profile and geometrybetween welds in a layer are evaluated; significantdeviations and excessive bumpiness are consideredindicative of a problem, and welding will be pausedfor the operator to evaluate the situation.

Figure 2 shows the operator interface for inter-acting with the weld plans. Bead cross-sectionestimations are placed inside a cross-section of theempty joint to be welded. Initially, the automatedweld planning algorithm displays the planned weld-ing sequence, but prior to execution the operator

Fig. 3. Experimental testbed (Liburdi F-Head with sensor bracket): (a) modeled sensor placements and fields of functional view; (b) actualsystem. Sensors (clockwise from bottom left): FLIR thermal camera, CCD camera, microphone, Basler CCD camera, Keyence LJV laser linescanner.

Piping and Pressure Vessel Welding Automation through Adaptive Planning and Control 529

will have complete freedom to make modificationswith intuitive, point-and-click-style interactions.Once welding begins, the plan is locked, and furthermanual modifications are halted unless anomaliesrequiring operator intervention are encountered.

Near-Real-Time Weld Defect Detection

One of the biggest cost-saving opportunitiesoffered by full welding automation is near-real-timedetection of possible weld defects. Early detectionwill allow defects to be addressed as they occur,making repairs and rework much simpler. Thecontrol system will utilize multisensor data fusionto predict the likelihood that one or more commonkinds of welding defects have occurred. To begin,four common defect types will be targeted. Table Ipresents the defects that this system aims to be ableto reliably detect. Different defect types can becaused by a variety of factors from improper powersettings to a dirty welding environment; however,based on literature concerning the welding processand time spent with highly qualified expert welders,an understanding of which physical features andprocess signatures characterize welding defects hasbeen developed;14–16 For example, it is commonknowledge among welders that oxides and contam-inants can form on the welds’ surface and promoteformation of internal voids, lack of fusion, andporosity between the layers. Therefore, carefulimage analysis of the weld layer surface prior toadditional welding should be able to identify poten-tial trouble areas, as they will have a different colorand reflectivity than the welded metal (Fig. 6; notethe strip of surface contaminant on the left-handside of the leftmost weld).

For algorithm development, welding tests will beconducted, and the process monitored and recordedby a sensor suite consisting of two CCD cameras, athermal camera, a microphone, and a laser linescanner (Fig. 3; Table I). Additionally, current and

voltage data from the arc will be recorded. Thethermal camera will be used to trigger interpasswait temperature commands as well as to evaluatethe thermal properties of the molten weld pool. Thisis expected to be useful in the detection of void andinternal porosity formation, as cold spots shouldappear in the liquid and solidifying metal. Themicrophone will be used to correlate acoustic emis-sions with the amount of metal (resistance) in thearc circuit. This is useful for both penetrationcontrol and determining void formation when pock-ets change the amount of metal in the circuit. TheCCD cameras will be used for judging and control-ling torch placement and weld overlap.

All sensed data will be recorded and analyzed tocreate the control system’s understanding of thenominal (good) welding process. Simultaneously,welding tests during which defect-causing condi-tions will be intentionally introduced will be con-ducted. The data resulting from these tests will beused to prove and refine the defect predictionalgorithms that will be based on a physical under-standing of how certain defects form. Multisensordata fusion algorithms will play a large role indeveloping quick prediction algorithms. Image andaudio processing routines will be combined withsensor fusion methods to predict defect formationnear the time of occurrence and define the necessaryadjustments in the process controls to return to anoptimal condition; For example, situations causinginternal voids will result in the weld metal notfilling in appropriately; a higher surface level(localized) would be expected on top of the void,since that metal has not fused down into the basemetal as desired. Additionally, work has been doneto correlate arc current and acoustic emission datawith the depth of metal under the electrode.5

Therefore, image processing to evaluate weld heightchanges can be paired with acoustic and currentdata processing to predict the likelihood of a voidhaving formed based on the in situ data collection. A

Table I. Defect types (see Fig. 3 for sensor names and placements)

Defect Cause Signature Correction Sensor/Fusion

(Potential)

Detection Methods

Surfaceporosity

Surface contaminants(soot, dirt, etc.), gas flow

Surface bub-ble/cavity; gas

sputter

Adjust gas flow;back up and remelt

Visible cam-era, (micro-

phone)

Surface texture andimage analysis

Void for-mation

Surface contaminants,poor placement, material

inconsistencies

Raised surface;cold spots; acous-

tic changes

Remelt; manualgrinding and auto-

matic reweld

(FLIR, visiblecamera,

microphone)

Acoustic/electricchange correlationsand surface analysis

Poor tiein

Poor placement Saw-toothededges; gaps

Shift torch/wire feedorientation; remelt;

filler weld

Visible cam-era, Keyence

Edge detection anddifferential/frequency

analysisInternalporosity

Surface contaminants,material inconsistencies,

gas flow

Gas sputter; coldspots; acoustic

changes

Remelt; manualgrinding and auto-

matic reweld

(FLIR, micro-phone)

Acoustic/electricchange correlationsand surface analysis

Robertson, Penney, McNeil, Hamel, Gandy, Frederick, and Tatman530

main question that this research will aim to answeris the resolution of the sensing system with respectto defect formation and detection. While the mech-anisms of formation are known, it is as of yetunknown how small of a defect can be detected andat what speed. These concerns are primarily whatwill be examined and addressed.

Initially, the focus will be on poor tie-in andsurface-level porosity, since those will be the easiestto predict or detect due to their surface-levelvisibility. Poor weld placement is a major factorfor tie-in issues and can be partially corrected at theplanning stage. Weld beads in the planning stagewill be placed to ensure good wetting on the sidewalls and overlap between beads, so that goodfusion is likely. However, even with good planning,real-world deviations can cause planning to goawry. For the in situ tie-in control, the front-facingcamera and the rear camera will be utilized. Thefront-facing camera is aimed at the front of the weldpool where the wire feed enters. This camera angleprovides the ability to view the edges of the liquidweld pool. Crowning on the cold edge of the weldbead is indicative of lack of fusion. Additionally, asaw-toothed edge and/or pockets on the weld bound-ary are indicative of poor fusion and can be identi-fied through image processing of either the rearCCD camera or potentially the rear-facing thermalcamera. The image processing to be used will besimple filtering and edge detection to determinewhether the weld boundary has deviated from asmooth, straight line along the interface of thecurrent weld and surrounding metal. Under anyscenario where poor fusion is detected, the controlmeasure will be for the torch to steer closer to thefusion boundary, which will promote tie-in at adistance based on the width of the defect area(Fig. 6).

Out-of-Position Aspects of Orbital Welding

When performing circumferential welds, as is thecase for actual industrial fabrication, gravity effectson the liquid weld pool will pose a challenge forrobotic control. To ensure that the welds stayconsistent as the welding head moves out of positionwith respect to the gravitational vector, a weld poolcontrol algorithm is being developed. A somewhatsimilar control algorithm is being developed using arobotic gas metal arc welding (GMAW) system inanother research project within this laboratory.Even though the welding process is different thanthe TIG system implemented in the groove weldingsetting, the dynamics of the weld pool are similar. Inboth cases, a pool of molten metal is in a situationwhere gravity can influence the flow of the liquidmetal and cause the weld pool to change to anundesired shape. This then would result in amisshapen weld bead that would cause defects inthe groove. The gravitational effects on the weldpool are typically manifested through a change in

geometry with a shift in the direction of gravity(Fig. 4). To combat this change in geometry, thereare several different control actions that can beused, e.g., regarding the work angle, the torchangle, and the position of the wire feed with respectto the weld pool. The work angle and torch angle, aswell as oscillation and travel speed, allow formanipulation of the orientation of the arc forcesagainst the weld pool, which are corrective mea-sures utilized by skilled human welders as well. Bychanging how the arc forces are applied to the weldpool, the gravitational forces on the weld pool can bemitigated. This can also be seen in Fig. 4, as thebeads welded with the torch oriented perpendicularto the base plate retain a more similar shape to thecontrol bead than the beads welded with the torchaligned with the gravitational direction. The posi-tion of the wire feed entry into the weld pool can alsoassist with more localized control of the weld pool bylowering the temperature of the weld pool at theentry point and thus increasing the surface tensionforces in the area around the wire. By increasing thesurface tension forces, the resistance of the weldpool to gravitational forces increases, thus helpingthe weld pool maintain the desired shape.

Further experiments have been run using theGMAW system to test how gravity affects the weldpool in the multilayered case. As can be seen inFig. 5, as layers continue to be stacked, the gravityeffects become more pronounced due to the increasein total heat in the part, as well as the shape of theprevious layers. In building the part shown inFig. 5, the torch was kept perpendicular to the baseplate, since this has been shown to produce a better-shaped weld bead than when the torch is alignedwith the gravitational direction (Fig. 4). The baseplate is angled at 45�, which is more than for theprevious welds shown, thus the gravitational effectson the weld pool are stronger. Using these results,further control algorithms are being developed touse feedback from cameras observing the weld poolas well as the known position of the robot to identifywhen the weld pool is drooping and reorient thetorch in a feedback loop that seeks to maintain thedesired, ideal weld pool geometry. The resultant

Fig. 4. Comparison of MIG welds at different base plate orientations.The leftmost bead is a control, welded at an angle of 0�. For thewelds in the top row, the welding torch was kept perpendicular to thebase plate, while for the welds in the bottom row, the welding torchwas kept vertical, aligned with the gravitational direction.

Piping and Pressure Vessel Welding Automation through Adaptive Planning and Control 531

control algorithm will then be updated to use thecameras specific to the TIG groove welding systemsuch that the added control points of the work angleand the wire feed position can also be used tofurther control the weld pool to produce the desiredbead shape in the out-of-position case.

EXPERIMENTAL SYSTEM DESCRIPTION

Figure 3 shows the experimental system aftermodifications. The welding system being used forexperimental tests and validation of the controlsystem is a Liburdi Dimetrics F style welding headconnected to a water-cooled P300 power system.This orbital welding system is designed for mecha-nized welding not computerized position control,thus significant modifications to the various motordrives were introduced into the system. Thesemodifications included adding servo actuators andcontrol for six axes on the Liburdi and adding asensor suite mounting bracket.

The sensor mounting bracket has been throughseveral iterations and currently houses a rear-facing Intertest iShot CCD camera (1280 9 720), arear-facing FLIR A35 thermal camera (320 9 256), afront-facing Basler ACE camera (659 9 494) with aneutral-density filter, a simple microphone mountedto the wire feed arm, and a Keyence LJV blue laserline scanner mounted in front of the welding torch.Since an ultimate goal of this research is commer-cialization and industrial deployment, all of thesystem modifications have been made as productagnostic as possible, and the sensors were chosenbased on both applicability and cost. A goal is toreduce the number of sensors needed once thecontrol algorithms are developed, to produce asminimal a sensor system as possible.

Several software development environments wereexplored for this research, and eventually NationalInstruments LabView software was chosen. It is acoding environment that is easy to use and isdesigned for sensor-based data acquisition andcontrol, which is a main objective of this research(Fig. 2). A major component of the software control

system under development is a graphical userinterface (GUI) for job planning and execution(Fig. 2c). This interface is used for manual andautomatic job planning, plan modification and visu-alization, and execution monitoring and control.

CURRENT RESULTS

Weld Planning and Adaptive Execution

Initial work has been performed to verify theexperimental system’s ability to reliably performwelds, as well as to evaluate the quality andcharacteristics of welds performed with minimal tono operator interaction. Dozens of welds wereperformed in single weld passes and multilayersequenced groove fills to evaluate the system’sperformance and note areas in need of enhancedcontrol, as well as identify strategies for control anddefect detection that became evident during thesetests.

After parameter tests with various sets of weldingparameters (current, voltage, travel speed, etc.)were performed, a specific set of parameters wasdecided on that produced consistent welds in linewith what would be desirable, as confirmed byskilled welding technicians and welding engineersat the Electric Power Research Institute (EPRI) inCharlotte, NC. The set of welding parameterschosen was a pulse weld oscillating between 190amps and 110 amps at 1.3 Hz, 10 V, 3 inches/mintravel speed, and a wire speed oscillating between22 inches/min and 18 inches/min (55.9 cm/min to45.7 cm/min) in conjunction with the current puls-ing. The groove and filler material are both stainlesssteel, and the filler wire is 0.045¢¢-diameter ER308L.The welds in Figs. 6 and 2d were performed withthese weld settings and are generally representa-tive of the shape, size, and surface of weldsproduced.

Using these welding parameters, two major setsof tests were performed. The first was a series of 20welds with the same parameters but incrementallydifferent placements laterally within the groove,

Fig. 5. Example of how gravity affects MIG weld beads in the context of additive manufacturing.

Robertson, Penney, McNeil, Hamel, Gandy, Frederick, and Tatman532

from the corner down the sloped walls to the center,and different torch angles towards the wall. Fourdifferent angles were used (0�; 3�; 6�; and 9�) withfive welding positions at each angle: two on thesloped wall, two on the flat center, and one in thecorner. These welds were performed to evaluatehow the weld surface and geometry change withvarying torch angle and as a function of locationwithin the groove geometry. Initial analysis indi-cates that the location in the groove has a moredominant effect on the resulting weld geometrythan the torch angle, as welds placed on the wallsdiffer greatly from welds placed in the corner orcenter of the groove, while welds performed in thesame position but with different torch angles arenearly indistinguishable. This is likely due to thevarying substrate geometry changing the way heattransfers from the weld pool to the substrate.Pending internal analysis, the torch angle isexpected to have a more noticeable role in thegeometric characteristics of the weld’s penetrationdue to a directional change of the arc forces.

The second was a series of multilayer, multiweldgroove fills using the same welding parameters asthe previous tests. During these, a series of weldswas planned and executed based on standardpractices demonstrated by welding technicians andengineers at EPRI. The welds were then executedautomatically by the system, with manual operatorintervention occurring only in situations where asystem error was encountered. These welds wereconducted to evaluate the system’s performancewithout adaptive planning adjustments, and toguide control system development. Errors andfaults during these tests also led to modifications

of the hardware and software system and thedevelopment of priorities in control system develop-ment. Typical welds can be seen in Fig. 6b.

Weld placement remained fairly constant as longas the base plate (substrate) alignment was main-tained, but obviously changed in cases where thebase plate came out of alignment. To facilitate baseplate movements, a robotic control concept calledoperational space control is employed. In thisapproach, control decisions must be made withrespect to the operational task space of the baseplate groove as opposed to the global coordinates ofthe robot system. Additionally, tests revealed thatthe wire feeder height is an important feature tocontrol, as variations in previous layer height andstop and start heights on previous weld slopes havea significant effect on what wire feeder height isideal at any given point. Sensing signatures becameevident in relation to this, as the behavior of thewire feeding into the weld pool generates significantand unique acoustic emission frequencies when toohigh above the weld pool, or grinding into the plateand shaking of the torch head when too low.

Finally, while few surface-level defects werefound, after cross-sectioning the welds, voids werediscovered in the corner toe of some of the initiallayers, indicating the likelihood of poor torch place-ment. This served to illustrate the need for bothprocess-dependent planning to avoid placementissues and in situ void detection to prevent fillingover problem areas sans repair.

Weld Defect Detection

An understanding of welding defects has beendeveloped such that defects can be produced asdesired for the purposes of study and detectionalgorithm development. Initial work has been doneto evaluate various image processing routines toidentify weld edges and individual welds. A majorfocus of upcoming experiments and research is thedevelopment and validation of defect predictionalgorithms. Table I presents the defect types thatdetection algorithms will be developed for, as well asthe sensor cues and sensing modalities that will beemployed for each of them.

As mentioned above, surface-visible defectsshould prove much easier to predict/detect in near-real time. Porosity is likely to have been caused bypoor gas coverage and can be detected visually aswell as audially. Since normal gas flow produces aconstant-frequency acoustic emission and volumet-ric rate, turbulent flow problems will be character-ized by a sputtering that will change the flowfrequency spectrum of the process in adetectable way. Additionally, pockets in the weldsurface will be identifiable via image processing ofthe solid welds from the rear-facing camera(s). Thepockets will both disrupt the smooth edges in anedge-fitting algorithm and differ significantly incolor from the surrounding metal, having a darker

Fig. 6. Resultant test welds. (a) Commonweld defects/indicators (saw-toothed edges for poor tie-in; bands of surface contaminants). (b) Asuccessfully filledgroove following theopen-loopautomated filling tests.

Piping and Pressure Vessel Welding Automation through Adaptive Planning and Control 533

pocket of color. In the event of porosity detection,the control system will query the weld power supplyto ensure that the gas flow is on and set to thecorrect rate; if it is not, then operations will beimmediately stopped, and the operator will benotified to check the gas supply and connections.

For tie-in control, the main issue is torch tipplacement and orientation. The two variables ofnote are the electrode tip position relative to thegroove wall (lateral position) and the angle of thetorch towards the base metal. Automatic voltagecontrol of the power supply maintains the torch tip’sposition relative to the base metal (vertical position)to ensure consistent voltage settings and will there-fore not be a control point for the adaptive controlsystem. Based on image processing and the torchposition, the edges and perimeter of the weld pooland beads will be identified. Smooth, linear edgesare indicative of good tie-in; therefore, similar toporosity, dark spots, holes, and nonlinear edgebehavior will be taken as indicative of poor tie-in.However, this should be fixable in situ withoutstopping operations unless a major inconsistencyhas developed. The control system, upon detectingpoor tie-in, will drive the lateral torch positiontowards the trouble area to close the gaps andsmooth out the edges. This motion will continueuntil smooth, consistent boundaries between thecurrent and surrounding welds stabilize.

If corrective measures for tie-in on one side of thegroove produce a potential tie-in issue on theopposite side of the groove, then the current weldwill be driven to a position that provides adequatespacing to ensure appropriate tie-in and fusion willoccur during the subsequent weld pass. The trajec-tory for the fill of the current weld will be updated to

its new position, and the planned welding sequencewill be modified for future reference. This will causethe planned operation to change based on theautomatic planning routine and will also initiatechanges to the weld model. The operator will benotified of the changes and given an option to pauseoperations.

Initial results for image analysis on welds withgood and bad tie-in can be seen in Fig. 7. Seen inboth Fig. 7a and 7b are the original image with edgedetection overlaid in red. Below the original imageis a graph showing the contour variance fit by theedge detection, and below that is the frequencyspectrum of the contour variance. As expected,edges with bad tie-in have increased low-frequencycontent, and generally more frequency content thanedges with good tie in, which should have relativelyconsistent, high-frequency content. This kind ofanalysis leads into automatic tie-in detection viaimage processing and edge analysis.

SUMMARY AND FUTURE WORK

The initial phase of this research was the conver-sion of the Liburdi welding head from open-loopvelocity control to servo position control and theintegration of a multisensor suite suitable for roboticsoperations. This phase also included the developmentof an appropriate operator interface and comprehen-sive data acquisition capability and is essentiallycomplete. A ‘‘point and click’’-style multibead/multi-layer weld planning interface based on a rudimen-tary weld bead model is now operational.

After the initial phase, work shifted to a focus onthe development and testing of adaptive weldplanning, weld defect detection, and non-gravity-

Fig. 7. Image processing and analysis of TIG welds with good and bad tie-in. (a) Jagged and inconsistent edges from poor tie-in (same welds asin Fig. 6a). (b) Straight, even lines with good tie-in (same welds as in 6b but rotated 90�).

Robertson, Penney, McNeil, Hamel, Gandy, Frederick, and Tatman534

aligned weld pool management algorithms. Prelim-inary weld tie-in problem detection based on weldsurface image processing is showing real promise.Separate research into large-scale additive metalsmanufacturing based on MIG has provided usefulinsights into weld pool molten metal behaviorassociated with non-gravity-aligned welding, whichalso occurs in orbital TIG welding.

The preliminary results obtained thus far areencouraging and fortify the belief that achievingrobust systems-level welding automation usingavailable sensing, control, and robotics technologiesis possible in the near future.

ACKNOWLEDGEMENTS

This research is being performed through the NSFIndustry/University Cooperative Research Centercalled the Materials and Manufacturing Joining andInnovation Center (Ma2JIC), which is led by the OhioState University. The specific industry Ma2JICmembers supporting this work are the Electric PowerResearch Institute (EPRI), Oak Ridge NationalLaboratory, and ITW/Miller Electric.

REFERENCES

1. A. Zalkind, Welding Shortage Fact Sheet. The AmericanWelding Society, Report, (2007).

2. J. Wilkey, Blog [Online] (2015). https://awo.aws.org/2015/11/the-future-remains-bright-for-skilled-welders/. Accessed 11Nov 2019.

3. S. Babu, Introduction to Integrated Weld Modeling, vol 22(ASM Handbook, Ohio, 2010).

4. S. Mishra and T. DebRoy, Tailoring gas tungsten arc weldgeometry using a genetic algorithm and a neural networktrained with convective heat flow calculations. Mater. Sci.Eng. A 454455, 477 (2007).

5. N. Lv, J. Zhong, H. Chen, T. Lin, and S. Chen, Real-timecontrol of welding penetration during robotic GTAWdynamical process by audio sensing of arc length. Int. J.Adv. Manuf. Technol. 74(1–4), 235 (2014).

6. S.B. Chen and N. Lv, Research evolution on intelligentizedtechnologies for arc welding process. J. Manuf. Process.16(1), 109 (2014).

7. J.P. Steele, C. Mnich, C. Debrunner, T. Vincent, and S. Liu,Development of closed-loop control of robotic welding pro-cesses. Ind. Robot Int. J. 32(4), 350 (2005).

8. ASME, ASME boiler and pressure vessel code, Standard(2017).

9. R.D. Conroy, Orbital narrow-gap GTAW process is gettingnoticed. Weld. J. 72(12), 15 (1993).

10. B.K. Henon, Orbital GTAW boosts production of leak-freetubing. Weld. J. 85(6), 56 (2006).

11. D. Baek, H.S. Moon, and S.H. Park, Development of anautomatic orbital welding system with robust weavingwidth control and a seam-tracking function for narrowgrooves. Int. J. Adv. Manuf. Technol. 93(1–4), 767 (2017).

12. J.K. Tatman, L.S. McCracken, and T.G. Hicks, Developmentof New Weld Heat Input and Dilution Equations for GasTungsten Arc Welding—Part 1, Conference Paper (2013).

13. J. K. Tatman, Development of improved equations for weldheat input and dilution control—part 2, conference paper(2018).

14. M. Stadler, P. Freton, and J.-J. Gonzalez, Influence ofwelding parameters on the weld pool dimensions and shapein a TIG configuration. Appl. Sci. 7(4), 373 (2017).

15. R Komanduri and Z. Hou, Thermal analysis of the arcwelding process: Part I general solutions, Metall. Mater.Trans. B 31(6), 1353–1370 (2000).

16. Y. Li, Y.F. Li, Q.L. Wang, D. Xu, and M. Tan, Measurementand defect detection of the weld bead based on online visioninspection. IEEE Trans. Instrum. Meas. 59(7), 1841–1849(2010).

Publisher’s Note Springer Nature remains neutral with re-gard to jurisdictional claims in published maps and institutionalaffiliations.

Piping and Pressure Vessel Welding Automation through Adaptive Planning and Control 535