Development of Tensile Shear Strength Prediction Model for GMAW Welds Using Laser Vision Sensor and Gaussian Process Regression

Article information

J Weld Join. 2025;43(4):356-363

Publication date (electronic) : 2025 August 31

doi : https://doi.org/10.5781/JWJ.2025.43.4.2

Sol Mi Lee ^*^,^**

, Dong-Yoon Kim ^*

, Jong-Kyu Park ^***

, Daewon Kim ^****

, Jiyoung Yu ^*^,

, Seung Hwan Lee ^**^,

* Flexible Manufacturing R&D Department, Korea Institute of Industrial Technology, Incheon, 21999, Korea

** Department of Mechanical Convergence Engineering, Hanyang University, Seoul, 04763, Korea

*** Advance Research Team, Hwashin, Yeongcheon, 38828, Korea

**** Chassis Materials Development Team, Hyundai Motor Company, Hwaseong, 18280, Korea

†Corresponding author: †willow@kitech.re.kr,

††seunghlee@hanyang.ac.kr

Received 2025 July 1; Revised 2025 July 7; Accepted 2025 July 15.

Abstract

This study developed a model to predict the tensile shear strength of Gas Metal Arc Welding (GMAW) welds using statistical analysis of laser vision sensor profiles and Gaussian Process Regression (GPR). Despite GMAW’s widespread use, welding defects due to disturbances like thermal deformation can compromise mechanical properties. To address this, weld bead external profile data, acquired via a laser vision sensor, and process variables were utilized for model development. A cumulative sum (CUSUM)-based change point detection algorithm extracted top and bottom plate boundary points rapidly and accurately from weld bead profiles. These points were transformed to minimize measurement condition influence. Subsequently, a GPR model was constructed, employing these transformed feature points and process variables as inputs, to predict tensile shear strength. The model demonstrated excellent predictive performance with an R² of 0.9587, RMSE of 13.9447 MPa, and MAPE of 8.6958 %. Analysis revealed voltage setting was the most influential variable in predicting tensile shear strength. Transformed feature point coordinates, representing the distance from the bottom plate boundary to the weld reinforcement feature point, also showed significant influence. This study confirmed that the tensile shear strength of GMAW weldments can be predicted accurately with limited input data and fast processing time.

Keywords: Cumulative sum algorithm; Gas metal arc welding; Gaussian process regression; Laser vision sensor; Machine learning; Tensile shear strength

1. Introduction

Gas metal arc welding (GMAW) is a widely utilized welding technique applicable to a range of materials, including steel and aluminum alloys. Its high welding efficiency and superior productivity make it a preferred choice across numerous manufacturing industries. The automation of the GMAW process has gained significant momentum with the integration of robotic systems and other automated technologies. Despite these advancements, existing automation solutions remain limited in their ability to manage a range of quality challenges, such as defects caused by thermal deformation, assembly misalignments, irregular weld profiles, and the associated decline in weld strength. Consequently, there is an increasing demand for research and development focused on automated non-destructive testing (NDT) technologies aimed at overcoming the current limitations in weld quality assessment within the GMAW process.

Until now, welding quality inspection of GMAW has generally been performed through destructive tests such as weld cross-sectional profile tests and mechanical performance tests. However, destructive testing methods are inherently limited, as they do not allow for comprehensive inspection of the product1-3). To address these limitations, ongoing research is focused on developing non-destructive testing and automation technologies for assessing arc welding quality. Key areas of investigation include quality prediction based on welding process parameters, artificial intelligence-driven weld quality estimation, real-time process monitoring and prediction using arc sensors, and weld profile measurement and quality inspection employing vision-based sensor systems.

First, in the area of quality prediction based on welding process parameters, research has focused on predicting characteristics such as porosity, bead profile, and penetration depth of the weld zone using regression models built on process parameters including welding current, voltage, and travel speed. Additionally, some studies have employed design of experiments (DOE) methodologies to examine the influence of these parameters on tensile strength4-8). Although these investigations have provided valuable insights into the primary associations between process parameters and weld quality, they fall short in adequately representing the intricate interactions and nonlinear dynamics present among the parameters. Consequently, studies have employed a range of artificial intelligence techniques, such as fuzzy logic, neural networks, and random forest models, to forecast welding defects, process conditions, and tensile strength9-12). However, relying solely on process parameters poses limitations in accurately pinpointing detailed aspects of weld quality, such as the precise locations of defects. To complement this approach, quality monitoring studies have utilized arc sensors to measure current and voltage signals. These measurements enable the evaluation of arc stability, detection of porosity, and prediction of contact tip-to- workpiece distance (CTWD)13-16). However, the arc sensor-based approach may face practical limitations in industrial settings, as it necessitates real-time data acquisition during the welding process, which can be constrained by equipment capabilities or environmental conditions. To supplement these methods, non-contact measurement technology known as laser vision sensors (LVS) has gained increasing attention. LVS technology enables the inspection of weld profile both during and after welding, offering greater flexibility in data collection with minimal limitations related to timing or workspace. In a previous study involving LVS, multiple weld profile datasets were collected to extract profile features, which were then used as input for a convolutional neural network (CNN) to predict tensile strength17). When utilizing accumulated data, processing a large volume of weld profiles can lead to increased computational time for the model. Another study introduced a data simplification technique by identifying key features of the weld metal profile, using boundary points of the weld region extracted from the profile data18). Boundary point extraction was carried out by repeatedly applying gradient operations at varying pixel intervals. Although previous studies have demonstrated the potential of laser vision sensor-based weld quality inspection, there remains a lack of research on image processing algorithms and real-time weld quality prediction technologies suited for industrial applications.

This study developed a predictive model for the tensile shear strength of GMAW welds by combining weld appearance data captured using an LVS with relevant welding process parameters. A statistical feature-based algorithm was applied to the weld appearance profile to extract the boundary points of the weld metal, from which additional feature points representing key shape characteristics were derived. To reduce the impact of measurement conditions, the feature points were transformed before being utilized as input data. Lastly, a Gaussian process regression (GPR) model was established to estimate tensile shear strength by utilizing both the transformed feature points and process parameters as input variables.

2. Experimental Method

2.1 Welding Materials

The welding material employed was Al5083-O with thicknesses of 2.0 mm and 4.0 mm. Test specimens were prepared by cutting to dimensions of 150 × 150 mm. The welding wire utilized was AWS A5.10 ER5356 having a diameter of 1.2 mm. Table 1 presents the chemical composition and mechanical properties of the base welding material, while Table 2 details the chemical composition of the welding wire.

Table 1

Chemical composition and mechanical properties of Al5083-O

Table 2

Chemical composition of AWS A5.10 ER5356

2.2 Welding and Analysis Method

The welding power source employed was the Daihen Welbee W350, which operates using an alternating current (AC) pulse waveform. An AC pulse waveform is a welding control technique that alternates periodically between direct current electrode negative (DCEN) and direct current electrode positive (DCEP) polarities19). The proportion of time spent in DCEN relative to DCEP is referred to as the electrode negative (EN) ratio (%), which influences both the penetration depth and the amount of material deposited.

The joint was formed using an overlap configuration (see Fig. 1). To achieve various weld profiles, the EN ratio was increased from 5.0 (75 A / 19.0 V) to 9.5 (149 A / 21.9 V) m/min at 0%, and from 6.0 (81 A / 17.2 V) to 11.5 (148 A / 18.4 V) m/min at 15%, with increments of 0.5 m/min. Additionally, the welding speed was maintained at 40 cm/min, and the contact tip-to- workpiece distance (CTWD) was set to 15 mm. The work angle was maintained at 45°, the advancing angle at 10°, and the shielding gas consisted of 100% argon delivered at a flow rate of 15 L/min.

Fig. 1

Schematic design of welding joint preparation

LVS was utilized to acquire the weld appearance profile, with measurements conducted at a rate of 100 profiles/second. The LVS had a reference distance of 50 mm, with a resolution of 0.02 mm along both the X and Z axes.

Three tensile shear test specimens were prepared from each weld sample, following the ISO 4136 standard, to evaluate the mechanical performance of the weld (Fig. 2). The tensile tests were carried out using a 300 kN universal testing machine (AG-X, Shimadzu Co.) at a testing speed of 3 mm/min. To minimize the effect of rotational moment during tensile testing, a spacer matching the thickness of the machined tensile shear specimen was attached. Tensile shear strength was calculated based on a cross-sectional area obtained by multiplying the specimen’s width by the thickness of the top plate, which was 2.0 mm.

Fig. 2

The configuration of test specimen

3. Data Processing and Development of Predi- ction Model

3.1 Extraction of Weld Profile Features

In lap welds, the profile of the weld metal significantly influences the tensile shear strength of the weld. Fig. 3 presents a schematic of a lap joint, with feature points A, B, and C identified as key locations representing the profile of the lap weld17,20). A preprocessing procedure was applied to extract weld feature points from the weld profiles captured by the LVS. The region of interest was specified to eliminate extraneous areas, yielding profile data consisting of 1,400 points. To reduce noise, the data was downsampled by a factor of five, resulting in 280 data points, and the resolution along the x-direction was adjusted to 0.1 mm. Subsequently, recognizing that the cumulative sum (CUSUM) algorithm, which accumulates mean and variance, is sensitive to initial data, the boundary point (A) between the weld bead and the lower plate, as well as the boundary point (C) of the upper plate, were detected independently. CUSUM is an algorithm that detects shifts by continuously summing indicators that signal when observed values at particular time points diverge from the expected statistical distribution21).

Fig. 3

Feature point (A, B, C) in the bead profile

The boundary point (A) of the lower plate was identified using profile data moving forward from the upper plate to the lower plate, while the boundary point (C) of the upper plate was detected using profile data moving in the reverse direction, from the lower plate to the upper plate. The cumulative sum is determined using the mean, variance, and sensitivity. A point is designated as a feature point if it exceeds the specified threshold. The sensitivity values were set to 0.35 for detecting point A and 0.25 for point C. Fig. 4(a) and (b) display the forward and reverse profile data, respectively, while the corresponding CUSUM cumulative statistics are presented in Fig. 4(c) and (d). The point where the cumulative sum statistic exceeds zero is defined as the change index (k). In the forward profile data, the height at this change index k is denoted as Y_a, while the corresponding horizontal position, calculated by multiplying the index by 0.1, is designated as X_a. For the reverse data, the reference index of the change was calculated by subtracting the change index k from the total data length. Multiplying this converted index by 0.1 defined the position, while the profile height at this index was designated as Y_c. After determining the equation of the straight line connecting points A and C, the y-intercept of this line was incrementally adjusted upward until it intersected the weld bead, with the intersection point defined as point B. Since the coordinates of points A, B, and C may differ based on the defined region of interest, even for the same location, zero-point calibration was performed by adjusting points B and C relative to point A’s (x, y) coordinates. The corrected positions were labeled as B’(Xb', Yb') and C’(Xc', Yc') and served as input features for the model predicting tensile strength.

Fig. 4

Bead profile (a, b) and cumulative sum statistic analysis using the CUSUM algorithm (c, d)

3.2 Development of Tensile Shear Strength Prediction Model

This study utilized a GPR model to effectively represent intricate nonlinear patterns within the data. GPR is a Bayesian-based, nonparametric machine learning approach known for its ability to deliver high predictive accuracy, even when trained on relatively small datasets22).

The prediction model utilized as input the profile coordinates, Xb', Yb', Xc', Yc' representing weld joint feature points, along with welding process parameters such as wire feed rate (WFR), current, and voltage. The output variable was the peak tensile shear strength, measured in MPa. Prior to model training, the input and output datasets were normalized to mitigate the effects of varying data scales. This study’s GPR model was developed using a combination of scaled Matérn and noise kernel functions.

4. Experiment Results and Discussion

4.1 Extraction of Weld Profile Features

Using the laser vision sensor profiles gathered under diverse welding conditions, the boundary points (A and C) of the welding bead were identified through the application of the CUSUM algorithm outlined in Section 3.1. The algorithm took an average of 5.65 ms per profile to extract the boundary points. The accuracy of the feature points extracted through the algorithm was evaluated by computing the mean absolute error (MAE) and root mean square error (RMSE) between the detected and true feature points. Because the Y coordinate depends on the detected X coordinate, the evaluation was performed solely on the X coordinate. The error metrics for the extracted Xb' and Xc' coordinates are summarized in Table 3. Because Xc' is derived from the distance between points A and C, its error metrics were larger than those of Xb', as the individual prediction errors for each point accumulated.

Table 3

Error metrics of extracted feature points for X_b’ and X_c’

Occasional data loss or distortion was observed during measurements conducted with the LVS. Measurement errors were identified due to the constraints of the LVS when the boundary between the weld bead and lower plate was situated below the bead (Fig. 5). After excluding data from these specific conditions, the feature extraction results were assessed, yielding an MAE of 0.0825 mm and RMSE of 0.2030 mm for Xb', and an MAE of 0.2211 mm and RMSE of 0.3534 mm for Xc'. These results demonstrate that, under conditions where the weld bead profile is clearly defined, the automated feature point extraction algorithm developed in this study can efficiently and accurately extract the weld bead profile. However, in instances of data loss or severe distortion of the weld profile, it was necessary to perform preliminary classification and supplementary preprocessing to enable accurate extraction of the weld feature points.

Fig. 5

Cross section and LVS bead profile comparison

4.2 Prediction of Weld Tensile Strength

4.2.1 Tensile Testing in Relation to Variations in WFR and EN Ratio

Fig. 6 presents the tensile shear strength results of the weld joints as a function of variations in WFR and EN ratio, which are key process parameters in the GMAW process. Error bars were employed to illustrate the variation in tensile shear strength across the test specimens, thereby offering a visual representation of the distribution in the measurement data.

Fig. 6

Graph of tensile shear strength by wire feed rate

4.2.2 Prediction Using GPR

Tensile strength for the test dataset was predicted using the trained GPR model. A comparison of predicted versus actual tensile strength values, as well as the model’s performance indicators, is illustrated in Fig. 7. The coefficient of determination (R²) was calculated to be 0.9587, indicating that the developed model accounts for over 95% of the variability in tensile strength. The RMSE was 13.9447 MPa, demonstrating a low level of prediction error and confirming the high accuracy of the model’s predictions. Furthermore, the mean absolute percentage error (MAPE) was calculated to be 8.6958%, indicating that the predicted tensile shear strength deviated from the actual values by approximately 8% on average.

Fig. 7

Prediction of tensile shear strength

To assess the influence of each input variable on the GPR model’s tensile strength predictions, Shapley additive explanations (SHAP) analysis was performed, with the findings illustrated in Fig. 8. SHAP is an approach that quantifies the impact of each input variable on the model’s predictive results, where higher SHAP values correspond to a more significant contribution of the variable to the prediction23). The analysis results indicated that the weld profile feature points (Xb', Xc', Yc') played a significant role in predicting the model’s tensile strength. This is attributed to the fact that these points capture crucial geometric information directly associated with the effective cross-sectional area that bears the load in a fillet weld, thereby playing a vital role in determining weld strength. Of the process variables, current was found to be highly influential, given its central role in controlling heat input and shaping the ultimate weld bead profile. Conversely, WFR was identified as having the least significance among the process parameters. This is attributed to the strong positive correlation between wire feed speed and current, which likely limited its additional contribution due to multicollinearity among the parameters.

Fig. 8

SHAP value for tensile strength prediction

5. Conclusions

This study developed a model to predict the tensile shear strength of GMAW welds by statistically analyzing weld profiles obtained via an LVS and employing a GPR model, with subsequent validation of its predictive performance.

1) A change point detection algorithm based on the CUSUM method was implemented to efficiently and accurately extract the boundary points between the upper and lower plates from the weld joint profile. The algorithm extracted feature points at an average rate of 5.65 ms per profile, showcasing a processing speed significantly faster than manual measurement methods. Evaluation of feature extraction accuracy revealed that Xb' exhibited an MAE of 0.3955 mm and RMSE of 0.4735 mm, while Xc' demonstrated an MAE of 0.0825 mm and RMSE of 0.2030 mm. Under conditions where the bead shape was measured without distortion, the accuracy improved, with Xb' achieving an MAE of 0.0825 mm and RMSE of 0.2030 mm, and Xc' attaining an MAE of 0.2211 mm and RMSE of 0.3534 mm.

2) A GPR model was constructed to predict tensile shear strength, incorporating transformed profile feature points (Xb', Yb', Xc', Yc') and process variables (WFR, current, voltage) as input parameters. The developed model exhibited strong predictive capability, accounting for over 97% of the variability in tensile strength. The model achieved an RMSE of 13.94 MPa and MAPE of 8.69%, indicating a low prediction error relative to the actual values.

3) SHAP analysis of the input variables revealed that weld profile feature points exerted the most significant influence on tensile strength prediction, while welding current demonstrated the highest relative importance among the process parameters. The WFR exhibited low relative importance, likely due to multicollinearity with the welding current.

Acknowledgment

This study was supported by the Ministry of Trade, Industry and Energy through the Material Component Technology Development Project titled ‘Domestication of Bonding Equipment for Electric Vehicle Chassis and Battery Case Assembly and Development of Smart Bonding Line’ (Project No. 20022489).

References

1. Kang T. H, Yu J. Y, Kim Y. M, Hwang I. S, Lee S. H, Kim D. Y. Weldability Evaluation of GMAW and GTAW for Al-6.7 wt.% Mg Alloy. J. Weld. Join 39(5)2021;:471–479. https://doi.org/10.5781/JWJ.2021.39.5.2.

2. Liu Y, Wang W, Xie J, Sun S, Wang L, Qian Y, Meng Y, Wei Y. Microstructure and mechanical properties of aluminum 5083 weldments by gas tungsten arc and gas metal arc welding. Mater. Sci. Eng. A 5492012;:7–13. https://doi.org/10.1016/j.msea.2012.03.108.

3. Hasanniah A, Movahedi M.. Welding of Al-Mg aluminum alloy to aluminum clad steel sheet using pulsed gas tungsten arc process. J. Manuf. Processes 312018;:494–501. https://doi.org/10.1016/j.jmapro.2017.12.008.

4. Yu J. Y, Kim D. Y.. Effects of welding current and torch position parameters on minimizing the weld porosity of zinc-coated steel. Int. J. Adv. Manuf. Technol 95(1)2018;:551–567. https://doi.org/10.1007/s00170-017-1180-6.

5. Kamble A. G, Rao R. V.. Investigation on effects of parameters of GMAW process on bead geometry, hardness and microstructure of AISI 410 steel weldments. Adv. Mater. Process. Technol 8(3)2022;:2450–2464. https://doi.org/10.1080/2374068X.2021.1912537.

6. Chaudhari P. D, More N. N.. Effect of Welding Process Parameters On Tensile Strength. IOSR J. Eng 4(5)2014;:1–5. https://doi.org/10.9790/3021-04550105.

7. Ishak M, M. Noordin N. F, Shah L. H. Parametric studies on tensile strength in joining AA6061-T6 and AA7075-T6 by gas metal arc welding process. IOP Conf. Ser.: Mater. Sci. Eng 100(1)2015;:012042. https://dx.doi.org/10.1088/1757-899X/100/1/012042.

8. Tesfaye F. Parameter optimizations of GMAW process for dissimilar steel welding. Int. J. Adv. Manuf. Technol 1262023;:4513–4520. https://doi.org/10.1007/s00170-023-11356-7.

9. Wu C. S, Polte T, Rehfeldt D. A fuzzy logic system for process monitoring and quality evaluation in GMAW. Weld. J 80(2)2001;:33–S.

10. Tafarroj M. M, Kolahan F.. A comparative study on the performance of artificial neural networks and regression models in modeling the heat source model parameters in GTA welding. Fusion Eng. Des 1312018;:111–118. https://doi.org/10.1016/j.fusengdes.2018.04.083.

11. Cho J. H. Prediction of Arc Welding Quality through Artificial Neural Network. J. Korea. Weld. Join. Soc 31(3)2013;:44–48. https://doi.org/10.5781/KWJS.2013.31.3.44.

12. Karthekeyan P. B, Pandiarajan N, Ranjit R, Krish- nankutty P, N. Mohamed M. R, Pandiarajan B. Tensile strength prediction in monel 400 weldments using classification and regression algorithms in machine learning. Mater. Res. Express 11(10)2024;:106520. https://doi.org/10.1088/2053-1591/ad87b1.

13. Lee D. W, Jin C. N, Rhee S. H. Study on Real- Time Porosity Defect Detection Through Neural Network Structure Optimization using Genetic Algorithm in GMAW. J. Weld. Join 39(5)2021;:542–551. https://doi.org/10.5781/JWJ.2021.39.5.11.

14. Yoon J. Y, Lee Y. M, Shin S. C, Choi H. W. A Study on Welding Process Algorithm through Real- time Current Waveform Analysis. J. Weld. Join 33(4)2015;:24–29. https://doi.org/10.5781/JWJ.2015.33.4.24.

15. Ju W. H, Ryu H. C, Lim K. S, Lee J. J, Park Y. H, Cho S. M. A Study on the Defect Detection Algorithm by Interval Statistical Processing Method of Arc Welding Waveform. J. Weld. Join 39(1)2021;:74–80. https://doi.org/10.5781/JWJ.2021.39.1.9.

16. Seo B. W, Jeong Y. C, Cho Y. T. Machine Learning for Prediction of Arc Length for Seam Tracking in Tandem Welding. J. Weld. Join 38(3)2020;:241–247. https://doi.org/10.5781/JWJ.2020.38.3.2.

17. Lee C. H, Kim D. Y, Cheon J, Yu J. Y. Prediction Model for Tensile Shear Strength of Gas Metal Arc Weld using a Laser Vision Sensor. J. Weld. Join 41(5)2023;:349–357. https://doi.org/10.5781/JWJ.2023.41.5.5.

18. Lee K. D, Hwang I. S, Kim Y. M, Lee H. J, Kang M. J, Yu J. Y. Real-Time Weld Quality Prediction Using a Laser Vision Sensor in a Lap Fillet Joint during Gas Metal Arc Welding. Sensors 20(6)2020;:1625. https://doi.org/10.3390/s20061625.

19. Park H. J, Kim D. C, Kang M. J, Rhee S. H. The arc phenomenon by the characteristic of EN ratio in AC pulse GMAW. Int. J. Adv. Manuf. Technol 66(5)2013;:867–875. https://doi.org/10.1007/s00170-012-4371-1.

20. Kim D. Y, Hwang J. H, Kim G. G, Kim Y. M, Yu J. Y, Park J. H. Prediction of Weld Tensile-Shear Strength using ANN Based on the Weld Shape in Aluminum Alloy GMAW. J. Weld. Join 41(1)2023;:17–27. https://doi.org/10.5781/JWJ.2023.41.1.2.

21. Page E. S. Continuous Inspection Schemes. Biometrika 41(1/2)1954;:100–115. https://doi.org/10.2307/2333009.

22. Rasmussen C. E. Gaussian processes in machine learning, in Summer school on machine learning. Springer, Berlin, German 2003;:63–71. https://doi.org/10.1007/978-3-540-28650-9_4.

23. Lundberg S, Lee S. I.. A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Pro- cessing Systems. Long Beach, USA 2017;:4768–4777. https://doi.org/10.48550/arXiv.1705.07874.

Article information Continued

This Journal follow the http://creativecommons.org/licenses/by-nc/4.0/ Creative Commons Attribution Non-Commercial No Derivatives License

Chemical composition (wt.%)									Mechanical properties
Si	Fe	Cu	Mn	Mg	Cr	Zn	Ti	Al	T.S (MPa)	Y.S (MPa)
0.40	0.40	0.10	0.40-0.10	4.57	0.05-0.25	0.25	0.15	Bal.	305	151

	MAE (mm)	RMSE (mm)
X’_b	0.1970	0.3152
X’_c	0.3935	0.4735