7/30/2019 06140790

1/4

Abstract This paper proposes a study on

monitoring of molten pool image during pipe

welding in gas metal arc welding (GMAW) using

machine vision. As circumferential butt-weldedpipes are frequently used in power stations, offshore

structures, and process industries, it is requires

advance technique of welding monitoring process.

The research was conducted for welding of mild

steel with controlled welding speed and

Charge-couple Device (CCD) camera as vision

sensor. The systems used GMAW machine with DC

current. Neural network used for simulation of weld

bead width. The experimental result shows the

effectiveness of the image processing algorithm and

simulation process.

I. INTRODUCTIONRC welding process is nonlinear and

multivariable-coupled, because it involves many

uncertainties, such as, influences of metallurgy, heat

transfer, chemical reaction, arc physics, and

magnetization [1]. Therefore, intelligent control

systems have been developed for modeling and

controlling the welding process, as they derive the

control performance based on human experience,

knowledge, and logic techniques, instead of

mathematical process models. Intelligent technologies

for robotic welding which contains computer vision

sensing, automatic programming for weld path and

technical parameters, guiding and tracking seam, expert

robot welding system, intelligent control of welding

pool dynamic and quality have been investigated [2-3].

Piping is a frequent structure in the constructions of

welding components such as for the petrochemical

industry, power plant, power plant components energy

storage, etc. Piping for nuclear and fossil power plants,

chemical plants, refineries, industrial plants, resource

recovery, and cogeneration units are most often shop

fabricated. Pipelines and other system involving long

runs of essential straight pipe sections welded together

are usually field assembled. In recent years, the use of

new welding processes, new alloys, fracture toughness

limitations and mandatory quality assurance (QA)

program have made piping fabrication and installationmuch more complex. Gas Metal Arc Welding

(GMAW) process is one of the commonly used welding

processes for fabrication of piping. Techniques using

added filler metal as inserts are effective in manual and

automatic applications. In automatic GMAW welding,

welding head orbits the weld joint on a guide track

motors and drive wheels need to move the head around

the track and a spool of filler wire. Oscillation and arc

energy are adjusted to permit greater dwell time and

heat input into the side walls [4].

Compared to plate welding, welding of pipes is more

difficult due to the characteristics of the weldingprocess. If the constant welding conditions are

maintained over the full joint length, the bead width

becomes wider as the circumferential welding of small

diameter pipes progresses. Therefore, the control of

bead width products has been very difficult to perform

by constant welding conditions. The automation of

bead width control requires the ability to adjust speed of

welding torch or control welding arc current.

This paper proposes the monitoring of molten pool

image during pipe welding using machine vision as

sensor. This system is designed in order to reduce

welding process complexity and process time. Thisexperiment is conducted by using common industrial

pipe that applied in the industry. With the proposed

system, monitoring process of pipe welding can be

performed. And neural network will simulate the result

of weld bead width.

Monitoring of Molten Pool Image during Pipe Welding

in Gas Metal Arc Welding (GMAW) Using Machine

VisionArio Sunar Baskoro1, Erwanto1, and Winarto2

1Laboratory of Manufacturing Technology and Automation, Department of Mechanical Engineering

Faculty of Engineering, Universitas Indonesia, Kampus UI Depok 164242Laboratoryy of Process Metallurgy, Department of Metallurgy and Materials Engineering Faculty of

Engineering, Universitas Indonesia, Kampus UI Depok 16424

E-mail: ario@eng.ui.ac.id, winarto@metal.ui.ac.id

A

ICACSIS 2011 ISBN: 978-979-1421-11-9

381

7/30/2019 06140790

2/4

II. WELDING SYSTEMThe pipe welding system developed in this study is

shown in Fig. 1. The system consists of a

circumferential welding manipulator, CCD camera, thepersonal computer, GMAW machine, microcontroller

and motor board to control stepper motors which is

used for the revolution. GMAW nozzle torch will be

held by manipulator and rotated along the welding railin 360o. The schematic of welding manipulator is shown

in Fig. 2.

Material used in this experiment was mild steel with

the diameter of pipe is 101.6 mm with 8 mm in

thickness. GMAW machine was DC type with CO2 as

shielding gas. Material properties and welding

conditions is shown in Table 1.

III. IMAGE PROCESSING ALGORITHM ANDNEURALNETWORKMODEL

A. Image ProcessingThe schematic of monitoring system is shown in Fig.

3. CCD camera captures the image in front of the torch.

Therefore, the molten pool will be captured and

processed using image processing algorithm. As shown

in Fig. 4, image of molten pool is the brightest part

compared other image. This molten pool image will be

analyzed and measured using image processing

algorithm and will be compared with the actual weld

bead.

The image processing algorithm is shown in Fig. 5.

After capturing the image of molten pool in 5.a, the

image is transformed into grayscale image, 5.b. After

thresholding process, 5.c, the width of molten pool is

determined as shown in Fig. 5.d.

Fig. 1. Schematic of welding system

Fig. 2. Schematic of welding manipulator

Fig. 3. Schematic of monitoring system

Fig. 4. Image of molten pool from CCD camera

TABLE1

MATERIAL PROPERTIES AND WELDING CONDITIONS

Base metal Mild steel

Diameter of pipe (mm) 101.6

Thickness of pipe (mm) 8

Density (g/cm3) 7.85

Melting point (oC) 1371-1454

Thermal conductivity (W/m.K at 25oC) 24.3-65.2

Welding machine DC

Electrode diameter (mm) 1.2

Nominal arc length (mm) 1.5Arc voltage, V (V) 18

Arc current, I (A) 180-220

Welding speed, v (cm/min) 5-10

Shielding gas CO2

Shielding gas, q (l/min) 15

ICACSIS 2011 ISBN: 978-979-1421-11-9

382

7/30/2019 06140790

3/4

This detection is started from vertical direction to

find upper and lower edge. Then the left and right side

of the edge will be detected and noted as width of

molten pool. This parameter is then calibrated with the

real weld bead width. 1 pixel is equal to 0.202 mm.

B.Neural Network Model

Neural network model is used for modeling the result

of welding process. The neural network model used in

this experiment is shown in Fig. 6. It consists of 5 units

in input layer which are rotation angle (), welding

velocity (v), width of molten pool resulted from image

processing (w), arc voltage (V) and arc current (I); 5

units hidden layer and 1 unit output layer which is

simulated width of molten pool (W). Training method

algorithm used Levenberg-Marquardt algorithm. Data

used in this training process were taken from the

welding result with the following parameters: arc

voltage was 18V, arc current were 180A, 200A, and

220A, and the number of sample for measurement is 42

samples per one time welding process.

IV. RESULTS AND DISCUSSIONFigure 8 shows the comparison result from actual

weld bead width and detected weld bead width from

image processing algorithm process. The result of weld

bead width has range from 6.6 mm 10.2 mm. The

average error from the detection process is 0.1 mm with

standard deviation is 1.3 mm. This error may occur

from the error of image processing algorithm such as

Fig. 5. Image processing result: (a) Original image of molten pool,

(b) Grayscale image, (c) Binary image, (d) Detected width of molten

pool

Fig. 6. Neural network model

Fig. 7 Appearance of weld bead

Fig. 8. Comparison result between actual width and detected width and its

error.

Fig. 9 Comparison result between of actual weld bead width and

simulation width calculated by neural network.

ICACSIS 2011 ISBN: 978-979-1421-11-9

383

7/30/2019 06140790

4/4

threshold value and error of edge detection.

Figure 9 shows the comparison result between actual

weld bead width and simulated weld bead width from

neural network. The average error from the simulated

weld bead width is 0 mm with standard deviation is 0.4

mm. This error may occur from the construction of

neural network and training parameters.From the experiment results, it shows the

effectiveness of image processing algorithm and neural

network simulation. The monitoring of molten pool

image during pipe welding using machine vision has

been conducted. And the neural network simulation to

model the weld bead width has been performed and the

simulation results conform to the experimental results.

The errors may come from the image processing

algorithm and neural network. By improving both

methods, the good performance of monitoring and

simulation can be achieved.

V. CONCLUSIONSMain results obtained by the investigation are

summarized as follows.

-An automatic welding process of pipe welding systemusing machine vision was constructed. CCD camera

captured the molten pool image and image processing

algorithm determined the width of molten pool from

the captured images. Neural network used for

simulating the weld bead width.

-The error from image processing algorithm fordetecting weld bead width is 0.1 mm with standard

deviation is 1.3 mm.-The error from neural network simulation to detect

weld bead width is 0 mm with standard deviation is

0.4 mm. It is shown that simulation results conform to

the experimental results.

-From the experiment result, it shows the effectivenessof image processing algorithm and simulation

process.

REFERENCES

[1] Saedi, H.R. and Unkel, W., Arc and Weld Pool Behavior forPulsed Current GTAW, Welding Journal, vol. 67 no. 11 (1998),

pp. 247-255.

[2] Chen, S.B, Qiu, T., Lin, T., Wu, L., Tian, J.S, Lv, W.X.,Zhang, Y., Intelligent Technologies for Robotic Welding,

Robotic Welding, Intelligence and Automation, LNCIS 299

(2004), pp. 123-143.

[3] Pires, J.N., Loureiro, A., Godinho, T., Ferreira, P., Fernando,B., Morgado, J., Welding Robots, IEEE Robotics and

Automation Magazine, June (2003), pp. 45-55.

[4] Nayyar, M.L., Piping Handbook, McGraw-Hill, SeventhEdition, US, 2000.

[5] Na, S.J., Lee, H.J., A Study on Parameter Optimization in theCircumferential GTA Welding of Aluminum Pipes Using a

Semi-Analytical Finite-Element Method, Journal of Material

Processing Technology, vol. 57 (1996), pp. 95102.[6] Kondoh, K., Ohji, T., and Ueda, K., Optimum Heat Input

Control in Arc Welding of Steel and Aluminum Pipe, Material

Transaction, JIM, vol. 39 no. 3 (1998), pp. 413419.

[7] Vilkas, E.P., Automation of the Gas Tungsten Arc Welding,Welding Journal, vol. 45 no. 5, (1966), pp. 410 - 416.

[8] Andersen, K. and Cook, G.E., Artificial Neural NetworksApplied to Arc Welding Process Modeling and Control, IEEE

transaction Industry Application, vol. 26 no. 9 (1990), pp.

824-830.

[9] Lim, T.G. and Cho, H.S., Estimation of Weld Pool Sizes inGMA Welding Process Using Neural Networks, Journal of

Systems and Control Engineering, vol. 207 no.1 (1993), pp.

15-26.

[10] Kanti, K.M. and Rao, P.S., Prediction of Bead Geometry inPulsed GMA Welding Using Back Propagation NeuralNetwork, Journal of Materials Processing Technology, vol. 200

(2008), pp. 300-305.

[11] Muramatsu, M., Suga, Y., Mori, K., Autonomous MobileRobot System for Monitoring and Control of Penetration

during Fixed Pipes Welding, JSME International Journal Series

A, vol. 46 no.3 (2003), pp. 391-397.

[12] Lee, C.Y., Tung, P.C. and Chu, W.H., Adaptive Fuzzy SlidingMode Control for an Automatic Arc Welding System,

International Journal of Advanced Manufacturing Technology,

vol. 29 (2006), pp. 481-489.

[13] Carrino, L., Natale, U., Nele, L., Sabatini, M.L., Sorrentino, L,A Neuro-Fuzzy Approach for Increasing Productivity in Gas

Metal Arc Welding Processes, International Journal of

Advanced Manufacturing Technology, vol. 32 (2007), pp.

459-467.

[14] Bae, K.Y., Lee, T.H., Ahn, K.C., An Optical Sensing Systemfor Seam Tracking and Weld Pool Control in Gas Metal ArcWelding of Steel Pipe, Journal of Materials Processing

Technology, vol. 120, issues 1-3 (2003), pp. 458-465.

[15] Di,L., Srikanthan, T., Chandel, R.S., Katsunori, I.,Neural-Network-Based Self-Organized Fuzzy Logic Control

for Arc Welding, Engineering Applications of Artificial

Intelligence 14 (2001), pp. 115-124.

[16] Brzakovic, D., Khani, D.T., Awad, B., A Vision System forMonitoring Weld Pool, Proceeding for the 1992 IEEE

International Conference on Robotics and Automation, Nice,

Frace-May (1992), pp. 1609-1614.

[17] Chen, S.B, Zhao, D.B., Lou, Y.J., and Wu, L., ComputerVision Sensing and Intelligent Control of Welding Pool

Dynamics, Robotic Welding, Intelligence and Automation,

LNCIS 299 (2004), pp. 25-55.

[18] Baskoro, A.S., Kabutomori, M., Suga.Y., Automatic WeldingSystem of Aluminum Pipe by Monitoring Backside Image of

Molten Pool Using Vision Sensor, Journal of Solid Mechanics

and Materials Engineering, vol. 2 no. 5 (2008), pp.582-592.

ICACSIS 2011 ISBN: 978-979-1421-11-9

384