<p>Gas metal arc welding (GMAW) is used across various industries due to its versatility. However, it carries significant health risks, particularly from the inhalation of fumes. While noise and radiation are also concerns, the focus on fume emission rates (FER) has been increasing in importance. Despite extensive research, the effects of various welding process variables on FER are not yet fully understood. Traditional methods for measuring these emissions are time-intensive and expensive, especially when considering the plethora of available materials and welding techniques. To address this topic, several FER models are being discussed in this paper, based on transient electrical and optical process data. For this task, an extensive database of FER measurements has been acquired with the respective electrical and optical time series, with 240 welds for model training and an additional 28 welds for validation. This database has been used to train different machine learning (ML) algorithms with varying transparency regarding the correlation between the transient process data and the FER. The investigation revealed that basic statistical modeling, using multiple linear regression, can achieve an average FER prediction accuracy of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${\pm 25}{\%}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mo>±</mo> <mn>25</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> across the FER operating range, measured by the average deviation of the predicted FER value from the measured value over the entire test data set. This applies to a broad range of wire feed rates, from <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${5}{\text {m} \, \text {min}^{-1}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>5</mn> <mrow> <mtext>m</mtext> <mspace width="0.166667em" /> <msup> <mtext>min</mtext> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> </mrow> </mrow> </math></EquationSource> </InlineEquation> up to <InlineEquation ID="IEq3"> <EquationSource Format="TEX">${12}{\text {m} \, \text {min}^{-1}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>12</mn> <mrow> <mtext>m</mtext> <mspace width="0.166667em" /> <msup> <mtext>min</mtext> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> </mrow> </mrow> </math></EquationSource> </InlineEquation>. Deep learning algorithms can improve the average FER prediction accuracy even further, reaching up to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">${\pm 12}{\%}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mo>±</mo> <mn>12</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> using neural networks. However, they may lack the correlation insights provided by other ML approaches. A correlation analysis of the data revealed that further investigations of different filler materials, etc., can be reduced to a few significant FER measurements. The implications of these findings are far-reaching. By incorporating these predictive models into modern welding equipment, we can create virtual FER sensors or observers on-site. This not only improves process efficiency but also reduces welders’ exposure to harmful fumes while enabling a more comprehensive analysis of FER within the broader manufacturing context.</p>

Data-driven modelling of the fume emission rate for gas metal arc welding

期刊论文
出版模式：

开放获取
发表日期：

2026年2月28日

查看全文

Dominik Scheurenberg,
Samuel Mann,
Benjamin Ebert,
Mark P. Sanders,
Mohamed Behery,
Gerhard Lakemeyer,
Robert H. Schmitt,
Rahul Sharma,
Sebastian Stemmler

摘要

Gas metal arc welding (GMAW) is used across various industries due to its versatility. However, it carries significant health risks, particularly from the inhalation of fumes. While noise and radiation are also concerns, the focus on fume emission rates (FER) has been increasing in importance. Despite extensive research, the effects of various welding process variables on FER are not yet fully understood. Traditional methods for measuring these emissions are time-intensive and expensive, especially when considering the plethora of available materials and welding techniques. To address this topic, several FER models are being discussed in this paper, based on transient electrical and optical process data. For this task, an extensive database of FER measurements has been acquired with the respective electrical and optical time series, with 240 welds for model training and an additional 28 welds for validation. This database has been used to train different machine learning (ML) algorithms with varying transparency regarding the correlation between the transient process data and the FER. The investigation revealed that basic statistical modeling, using multiple linear regression, can achieve an average FER prediction accuracy of ${\pm 25}{\%}$ $\pm 25 %$ across the FER operating range, measured by the average deviation of the predicted FER value from the measured value over the entire test data set. This applies to a broad range of wire feed rates, from ${5}{\text {m} \, \text {min}^{-1}}$ $5 m {min}^{- 1}$ up to ${12}{\text {m} \, \text {min}^{-1}}$ $12 m {min}^{- 1}$ . Deep learning algorithms can improve the average FER prediction accuracy even further, reaching up to ${\pm 12}{\%}$ $\pm 12 %$ using neural networks. However, they may lack the correlation insights provided by other ML approaches. A correlation analysis of the data revealed that further investigations of different filler materials, etc., can be reduced to a few significant FER measurements. The implications of these findings are far-reaching. By incorporating these predictive models into modern welding equipment, we can create virtual FER sensors or observers on-site. This not only improves process efficiency but also reduces welders’ exposure to harmful fumes while enabling a more comprehensive analysis of FER within the broader manufacturing context.