Estimation of Welding Current with Adaptive Neuro Fuzzy Inference System (ANFIS): Utilization of Arc Light Signal Emitted in the Arc Welding Process

Abstract

The main purpose of this study is to estimate the welding current using the arc light signal emitted during the welding process. Traditionally, welding operators determine this current from the arc light based on their visual perception. This study shows that, using artificial intelligence techniques, welding current can be automatically estimated through arc light and can also be useful for monitoring of the process and detecting its disturbances. For this purpose, initially, a data acquisition system is designed to synchronize the movement of the light sensor with the electrode holder. The electrode welding machine is set to different maximum current levels, and two electrodes with different diameters are used at each level. During the welding process, the arc light and current signals are acquired simultaneously. The obtained data are filtered and aligned by cross-correlation. For the ANFIS (adaptive neuro-fuzzy inference system) model, the arc light is defined as the input and the current as the output. The estimation results of ANFIS are further improved through filtering, shifting, and current-limiting processes. The maximum cross-correlation values for training and testing data are 0.9587, 0.9598, 0.9565, and 0.9323, respectively, while the R-squared values are 0.7033, 0.7640, 0.6449, and 0.5853. Compared with the artificial neural network (ANN) model, it is observed that the ANFIS model provides better prediction results. The results confirm that arc light signals can be effectively used for welding current prediction. Therefore, the proposed approach can contribute to the development of intelligent welding systems and quality welding processes by reducing operator dependency.

1. Introduction

Arc welding processes have great importance in industrial applications, and parameter estimation plays a critical role in welding automation. Arc light, one of the welding parameters, is a fundamental phenomenon associated with the arc welding process, and its emission provides valuable information about the welding process [1].

The spectral distribution and visible radiation of the arc light are used as an important indicator for monitoring the welding process and evaluating its quality [2,3,4,5]. Various techniques have been developed to obtain information from the arc light, and visual perception has the advantage of providing this information without requiring physical contact [6]. These techniques are divided into two categories as active and passive in the context of visual perception systems and are effectively used in the welding process [7]. In passive vision, the object is directly viewed, while in active vision, laser or structured light is reflected onto the object surface. Active and passive visual sensors (camera, luxmeter, laser vision systems (LVS), structured light sensors, etc.) are used together with artificial intelligence integration and intelligent welding robots to provide effective results in the welding process. These sensors play an important role in areas such as preventing welding noise, process control, and quality improvement [7,8,9,10,11]. For example, photodiodes and light intensity measurements have been used in the detection of the properties of weld melt pools and in obtaining shape information [6,12]. In addition to visual detection in welding processes, arc, acoustic, and other detection techniques (touch, electromagnetic, infrared, temperature, ultrasonic, radiographic) are widely used [13,14,15,16]. Studies in the literature show that the combined use of different detection methods in welding processes provides significant advantages in terms of process monitoring, algorithm development, fault detection, quality control, fusion algorithm, and methodology development [17,18,19,20].

Arc light is also closely related to various welding parameters. It has been determined that there is a functional relationship between arc light intensity and arc length [21] and between arc brightness and current [22]. In this context, to analyze the parameters related to arc light more accurately, it is necessary to eliminate unwanted data in the acquired signals. In this direction, filtering is a widely used method. Physical optical and digital filters are frequently preferred. In the study of da Silva et al. [22], arc brightness was measured with a luxmeter and protected using a level 6 welding mask light filter (physical optical filter). In order to obtain a more accurate digital signal by filtering unwanted frequency components, Wei et al. [23] passed the arc voltage and current signals through Butterworth low-pass filters (digital filters) before entering the data acquisition board (DAQ board). In addition, in another study by Jian Lin et al. [24], a Gaussian low-pass filter (digital filter) was applied to the arc signal obtained from the welding arc sensor.

The acquired and filtered signals are effectively used in control systems for parameter estimation. Closed-loop control systems developed using arc light [25] and welding current determination models proposed by Kim et al. [26] are important examples for the optimization of this process.

In this context, effective processing and interpretation of data obtained through process control enables the development of more intelligent and adaptable systems. Recently, machine learning and artificial intelligence techniques have become increasingly important in terms of creating data-driven decision mechanisms and parameter estimation in welding processes. Machine learning is an artificial intelligence application that enables systems to learn through experience and plays an important role in increasing efficiency in welding processes [27].

Among artificial intelligence methods, the neuro-fuzzy approach stands out as a hybrid system formed by combining neural networks and fuzzy logic [28]. Based on this approach, Jang [29] developed the ANFIS architecture using a hybrid learning procedure. In particular, the ability of ANFIS to model nonlinear functions in welding processes makes it a powerful tool in the field of parameter estimation [28]. ANFIS has been used as an effective method for analyzing the effects of welding parameters, making real-time predictions and automatically extracting weld seams [30,31]. In addition, ANFIS can be effectively used in predicting weld quality and in determining metal deposition and seam width [32,33,34,35]. Furthermore, ANFIS has been successfully applied in various fields such as analyzing the responses of human welders, estimating penetration, understanding surface reactions, and developing intelligent welding machines [36,37,38,39]. Another alternative artificial intelligence method frequently used in welding processes is ANN. Eguchi et al. [40] developed a neural network (NN) arc sensor that estimates the electrode wire and arc lengths using voltage and current measurements. In addition, ANN models have been effectively used to determine various weld properties such as quality, location, optimal working point, seam width, penetration depth, seam geometry, and the joining of dissimilar metals [41,42,43,44,45,46].

In recent studies, the prediction results of ANFIS and ANN models are directly or indirectly compared. Gokmen et al. [47] estimated the average power consumption of an electrode welding machine using ANFIS with noise characteristics and compared this result with ANN. Performance comparisons of ANFIS and ANN models in estimating weld seam width and penetration depth were made [48,49]. In the study conducted by Devaraj et al. [50], ANFIS and ANN models were also compared according to regression analysis to estimate weld properties such as dilution and weld seam geometry, and the results were evaluated according to root mean square error (RMSE). There are also studies in which different methods such as ANFIS and MGGP were examined comparatively in estimating welding processes [51]. These studies show that both ANFIS and ANN are effective methods for controlling, predicting, and optimizing the welding process.

In this framework, although the studies in the literature have presented various approaches for light and arc welding parameters of the welding process in artificial intelligence, the studies on the light and current estimations are limited. At the same time, current estimation with ANFIS light input is a subject that has not been fully elucidated in the current literature. This method aims to make current estimation with light signals, which has not been sufficiently studied in the scientific literature before. The results provide important findings on current estimation with arc light signals and contribute to research in this field. In conclusion, the integration of such artificial intelligence approaches into welding processes has the potential to contribute to increased efficiency, safety, and sustainability, not only in specific areas such as current estimation but also in broader areas such as improving weld quality, parameter optimization, and energy efficiency. Furthermore, it demonstrates that it can be useful for monitoring the course of the process, detecting disturbances and anomalies.

2. Materials and Methods

In this study, to better perceive the current change, the maximum current level of the electrode welding machine was set to 80 amperes (A), 100 A, 120 A, and 140 A, respectively, and arc light and current signals were acquired simultaneously with two different electrode types at each level. A system was designed to ensure the simultaneous movement of the electrode and light sensor; signal filtering and aligning processes were applied to the obtained data. ANFIS and ANN models were used to estimate current with arc light data and to compare the models. Similar arc brightness has been measured at different current levels by da Silva et al. [22].

2.1. Data Acquisition

To ensure that the welding electrode holder and the light sensor move simultaneously during the data acquisition process, the system shown in Figure 1a–d was designed. This system made it possible to acquire arc light data precisely.

The system design shown in Figure 1a,b configures the electrode holder and the light sensor to move simultaneously. This design was developed to ensure that the light sensor remained at a constant distance from the arc light emitted during the welding process and to optimize its movement in the three-dimensional plane. In addition, welding glass with number 9 was integrated into the system to filter the arc light data. The applied version of the design is shown in Figure 1c,d.

The data acquisition system and sequence are shown in Figure 2. The welding process was carried out using a Magmaweld Monotig 200i (Turkey) manual metal arc welding (MMAW) machine (1, 2). Arc light data were acquired using a CEM DT 1309 (China) model luxmeter (3) and CEM DT 1309 interface software (4, 5). Current data were acquired using a Lem LA 305-S (Swiss) current transducer (6), an NI USB-6001 (USA) DAQ card (7), and LabVIEW 2021software (8, 5). Necessary calibration procedures were performed for Lem transducer. All calculations, analyses, and artificial intelligence applications were performed using MATLAB software (R2024a). LabVIEW software (Version 21.0) was used for acquiring current data and performing calibration calculations. DT1309 interface program was used to acquire arc light data. The current and arc light signals were acquired at a sampling frequency of 9 Hz for synchronization between the CEM DT 1309 and NI USB-6001.

Examples of experimental applications of the study are presented in Figure 3.

2.2. Data Preprocessing

The acquired signals were filtered with a low-pass Butterworth filter to eliminate noise. Third degree, 50 Hz sampling frequency, and 5 Hz cut-off frequency parameters were used for filtering process. This step was applied to increase the accuracy of the signals and improve the data quality. In addition, shifting (shifting on the sample axis) process was performed with a cross-correlation method to find the alignment between two signals. This method was applied to evaluate the degree of overlap of signals, to facilitate analysis, and to ensure that artificial intelligence models produce more accurate results. Finally, the acquired signals were optimized by separating them into training and test data.

2.3. Adaptive Neuro Fuzzy Inference System (ANFIS)

ANFIS is a combination of ANN and fuzzy logic and is effective in modeling nonlinear functions and estimating chaotic time series [29]. ANFIS was used in various studies to estimate arc welding parameters [47]. In the study, ANFIS was used to make current estimation with arc light data. Arc light data were used as input and current data as output for training the models. To enhance the estimations of the ANFIS model, the estimated results were filtered using the same filtering method employed during preprocessing. Subsequently, cross-correlation analysis was conducted to align the filtered estimated signal with the shifted current signal. Minimum and maximum 0 A–190 A current limitation was applied to the outputs of models.

2.4. Artificial Neural Network (ANN)

ANN is a large-scale parallel distributed processor consisting of simple processing units that can store and make usable experiential knowledge [52] (p. 2). It has been shown in various studies that ANN is effective in processes such as welding monitoring, quality estimation, and electrode position [42,43].

In the study, hidden layers were added to the ANN. The multilayer ANN was trained to estimate current values using arc light data as input. The signal processing methods applied to the ANFIS model were also applied to the ANN.

2.5. Statistical Metrics

Cross-correlation consists of the shifted dot product between two series and is generally used to measure the degree of similarity between two signals. It is also a crucial tool for determining the time delay. In the case of discretized records, the cross-correlation between two signals u and v with the same N samples length is provided in Equation (1) [53]. Normalization was applied to the obtained cross-correlation results to ensure that the values were between “−1” and “1”. Cross-correlation analysis is applied to detect leaks in liquid-filled pipes [54]. The true R-square is the proportion of the variance in Y that is explained by X. The R-square and root mean square error (RMSE) formulas are presented in Equations (2) and (3) [55] (pp. 152–290). In the literature, R-squared has been used to evaluate the performance of machine learning models in predicting hydrogen production [56]. In addition, the ANFIS method was used to predict the heat-affected zone, surface roughness, and the results were evaluated using RMSE [51].

$$corr[u,v]\left(t\right)={\sum }_{n=n0}^N\left(u\left(n\right).v(t+n\right))$$

(1)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\mathrm{ŷ}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-{\mathrm{ȳ}}_i\right)}^2}}$$

(2)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-{\mathrm{ŷ}}_i\right)}^2}$$

(3)

Equation (1) expresses the cross-correlation between the time series u and v. In Equation (2), y_i represents the observations; ŷ_i represents the values predicted by the model; and ȳ_i represents the mean value of the observations. In Equation (3), y_i represents the observations, and ŷ_i represents the predicted values.

3. Results

In this study, the use of arc light signals to estimate the current signals in arc welding processes was investigated. Arc light and current data acquired in two different electrode types and four different current levels were filtered, and signal shifting was performed and analyzed with the ANFIS model. The current estimation performance of the ANFIS model was compared with the ANN model.

3.1. Data Analysis

In application, the arc light (lux, lx) and current signals were simultaneously acquired using rutile-coated electrodes with dimensions of 3.25 × 350 mm and 2.50 × 350 mm, with the electrode welding machine set to maximum current levels of 140 A, 120 A, 100 A, and 80 A, respectively. Datasets of different lengths were obtained for each amperage level. The obtained current and arc light datasets have been adjusted to equal values for signal analysis and estimation tasks.

Table 1. Raw, training and test data.

Current and Arc Light Data	Electrode (mm)	80 A	100 A	120 A	140 A
Raw data (arc light and current)	3.25	2320 × 1	1948 × 1	2014 × 1	1600 × 1
	2.50	1276 × 1	1116 × 1	945 × 1	919 × 1
Train data (arc light and current)	3.25	2041 × 1	1802 × 1	1835 × 1	1481 × 1
	2.50	1177 × 1	817 × 1	746 × 1	820 × 1
Test data (arc light and current)	3.25	279 × 1	146 × 1	179 × 1	119 × 1
	2.50	99 × 1	299 × 1	199 × 1	99 × 1

shows the raw, training, and test data separately. The current and arc light datasets were concatenated to create raw, training, and test datasets, and their sizes were equalized.

Figure 4a–d show the in situ measured current and arc light graphs during welding processes using electrodes with diameters of 3.25 mm and 2.50 mm, respectively. During the experiment, certain intervals were left as gaps in the welding process to better analyze the effects of electrode movements and to evaluate the artificial intelligence prediction process more effectively. These gaps are consistent with the current change periods seen in Figure 4a,c. The changes in the periods also reflect the dynamic movements of the welder’s electrode to some extent. When comparing the current levels in Figure 4a,c with the light levels in Figure 4b,d, it is observed that the arc light level increases as the current level increases. The arc light and welding current are sensitive to the movements of the operator. For instance, at a current level of 140 A in Figure 4b, the arc light intensity reaches 274.3 lx at the 294th sample and 186.4 lx at the 781st sample. Similarly, in Figure 4d, at the same current level, the arc light intensity reaches 344.6 lx at the 193rd sample and 290.3 lx at the 730th sample. This corresponds to instances where the welder exhibits unstable movements with the electrode or temporarily interrupts the welding process during a flawless welding process; when the appropriate current levels are chosen, the arc light follows a more stable trajectory. For example, at 120 amperes (3.25 mm electrode), the arc light intensity reaches a stable average level of approximately 25 lx, while at 100 amperes and 80 amperes, it stabilizes around 16 lx and 10 lx, respectively. In cutting regions, transition values are smoother, and sudden fluctuations are reduced. In Figure 4d, at high current levels using a 2.5 mm diameter electrode, the arc light intensity shows pronounced fluctuations due to errors in the welding process caused by mismatches between the electrode diameter and the current.

3.2. Data Preprocessing Results

In this study, the obtained data were processed by applying filtering, and the main components were preserved. In addition, arc light and current signal were shifted using the cross-correlation.

The graphs presented in Figure 5 show the merging of measured current and light data between 80 A and 140 A, respectively, according to the maximum current levels of the electrode welding machine. This process was performed to match the temporal shift between current and light signals during the welding process and to increase its usability in artificial intelligence prediction processes. In Figure 5a,c, filtered arc light signal (FALS) and filtered current signal (FCS) obtained at maximum current levels of 140 A, 120 A, 100 A, and 80 A during the welding process performed with 3.25 mm and 2.50 mm diameter electrodes are presented. In Figure 5b,d, the FALS and shifted filtered current signal (SFCS) obtained through the shifting process are presented.

3.3. ANFIS Performance Results

In this study, the ANFIS model was applied for current estimation. It was observed that ANFIS models have demonstrated effectiveness in various applications by combining the learning capabilities of ANN with the reasoning abilities of fuzzy logic [31,32,33,34,35,36,37,38].

Figure 6 and Figure 7 show the fuzzy inference system (FIS) model and the structure of ANFIS model, respectively. In Figure 7, the logical operations in the ANFIS model are shown as blue (AND operation), red (OR operation), and green (NOT operation). However, only the AND operation is used in the designed ANFIS model.

The parameters were determined as 30 training cycles, 200 membership functions, and “gbellmf”-type membership function in ANFIS model. A total of 200 ANFIS rules were created within the scope of the study.

Different experiments were performed for each amperage level (140 A, 120 A, 100 A and 80 A), and the data obtained were combined to form training and test sets. The training and test results of the experiments with electrodes of 3.25 mm in diameter are presented in Figure 8 and Figure 10, respectively; also, the same was done for the 2.50 mm diameter electrode shown in Figure 9 and Figure 11.

Analysis of the data presented in Figure 8a reveals a strong overall agreement between FSACO and the target FCS. However, the same figure also indicates a degree of lag and some deviations in FSACO’s response to abrupt changes in the FCS. A similar trend is observed in Figure 9a. The sharp peaks shown in the cross-correlation graphs in Figure 8b and Figure 9b indicate that FSACOs with diameters of 3.25 mm and 2.50 mm effectively follow the FCS.

As a result of the performance metrics calculations, regarding the data presented in Figure 8a, the R-squared value between the FSACO and the target FCS was calculated as 0.7033, and the RMSE value was 32.6174 A. According to the data in Figure 8b, the maximum cross-correlation value between the FSACO and the target FCS was determined as 0.9587, and the comparison of the FSACO and the FALS is presented in Figure 8c. Similarly, according to the data presented in Figure 9a, the R-squared value was calculated as 0.7640 and the RMSE value as 29.4357 A. In Figure 9b, the maximum cross-correlation value was found to be 0.9598. The data presented in Figure 9c show the relationship between FSACO and FALS.

In Figure 8c, it is observed that FSACO responds to the gradual increase and decrease behaviors of FALS, which are based on the movements of the welder, and follows FALS in harmony. However, in Figure 8a, it is observed that the welding current FCS cannot respond at the same speed to the sharp changes in FALS corresponding to the welder’s sudden movements. A similar trend is observed in Figure 9a. These results reveal that the ANFIS model can be successfully applied on the training data and accurately capture the real current signal.

In Figure 10 and Figure 11, it is observed that the models produce consistent estimations at consistent current levels. In Figure 10a, FSACO exhibits a more stable performance at high current levels, while significant deviations occur at low currents and current changes. Similarly, in Figure 11a, FSACO exhibits more deviations at low currents and current changes compared to high currents. The distinct but not sharp peaks in Figure 10b and Figure 11b indicate that FSACOs with diameters of 3.25 mm and 2.50 mm successfully track FCSs.

According to the data presented in Figure 10a, the R-squared value was calculated as 0.6449 and the RMSE value as 33.5493 A between the FSACO and target FCS. In Figure 10b, the maximum cross-correlation value was obtained as 0.9565 between the FSACO and target FCS. Figure 10c shows the relationship between the FSACO and FALS. Similarly, according to the data presented in Figure 11a, the R-squared value was calculated as 0.5853 and the RMSE value as 38.9470 A. In Figure 11b, the maximum cross-correlation value was found to be 0.9323. Figure 11c shows the relationship between the FSACO and FALS. These results verified that the ANFIS model is successful for making accurate estimates on test data at different current levels for both electrodes. In terms of performance metrics, the ANFIS model created for the 3.25 mm electrode based on the test data gave more successful results compared to the 2.50 mm electrode model.

3.4. ANN Performance Results

In this study, the ANN model, like the ANFIS model, was trained using arc light as input and current data as output. The estimation of the quality of welding processes has been performed using ANN models in several studies [40,41,42,43,44,45,46].

The designed ANN model structure is shown in Figure 12. The model consists of three hidden layers with 40, 30, and 20 neurons, respectively, and one output layer. The Levenberg–Marquardt backpropagation algorithm was used for training.

In Figure 13 and Figure 14, the ANN estimations made at current levels of 140 A, 120 A, 100 A, and 80 A using training data obtained from electrodes with diameters of 3.25 mm and 2.50 mm were compared with the target current signals.

According to the data in Figure 13a, the RMSE value between FSANNCO and the target FCS was found to be 33.6504 A, and the R-squared value was 0.6842. In Figure 13b, the maximum cross-correlation value between FSANNCO and the target FCS was found to be 0.9559. According to the data in Figure 14a, the R-squared value was 0.7364, and the RMSE value was 31.1072 A. In Figure 14b, the maximum cross-correlation value was calculated to be 0.9550.

The R-squared value between FSANNCO and the target FCS was calculated as 0.6417 and 0.5412, respectively, for the 3.25 and 2.50 mm electrode ANN model test data. The RMSE was found to be 33.7037 A and 40.9669 A, respectively. The maximum cross-correlation value was found to be 0.9554 and 0.9349, respectively.

3.5. Comparison of ANFIS and ANN Performance Results

To evaluate the performance of the ANFIS model, a performance comparison with the ANN model was carried out. Studies on the comparison of the performances of ANFIS and ANN models, model outputs, and errors were shown in [47,49]. In the literature, studies have been conducted comparing the performances, model outputs, and errors of ANFIS and ANN models [47,48,49,50]. The estimation results and performance metrics of ANN and ANFIS models are presented in

Table 2. Training and testing of results of ANFIS and ANN.

Model	Phase	R-Squared	Cross-Correlation	RMSE (A)
ANFIS 3.25	Train	0.7033	0.9587	32.6174
ANFIS 3.25	Test	0.6449	0.9565	33.5493
ANFIS 2.50	Train	0.7640	0.9598	29.4357
ANFIS 2.50	Test	0.5853	0.9323	38.9470
ANN 3.25	Train	0.6842	0.9559	33.6504
ANN 3.25	Test	0.6417	0.9554	33.7037
ANN 2.50	Train	0.7364	0.9550	31.1072
ANN 2.50	Test	0.5412	0.9349	40.9669

.

When the performance metrics of the ANFIS and ANN models were examined, it was observed that the ANFIS estimation results were more consistent compared to the ANN model. However, in the 2.50 Cross-Correlation test result in Table 2, it was determined that the ANN provided better results than the ANFIS.

The comparison of FSANNCO and FSACO model results with the target FCS data, using the training data for the 3.25 mm electrode and 2.50 mm electrode, is shown in Figure 15 and Figure 16. In Figure 15 and Figure 16, although the tracking trends of FSACO and FSANNCO with the target FCS show similarities, it is observed that FSACO is more successful in capturing FCS changes. As shown in Figure 15 and Figure 16, it has been observed that FSACO produces more stable results, particularly at low current levels and during sudden current changes. Additionally, it has been observed that FSACO is more effective in capturing dynamic changes in the welding arc during transitions between different current levels compared to FSANNCO.

In Figure 17a,b, the test data results of the ANN model for 2.50 and 3.25 mm electrodes are shown and compared with FSANNCO and the target FCS.

Figure 17a shows that although FSANNCO tracks the FCS effectively, it experiences deviations in the transition regions. In contrast, Figure 17b reveals a more stable performance of FSANNCO in tracking the FCS graph.

When the performance metrics and graphical analyses of ANFIS and ANN models are examined, it is observed that ANFIS achieves higher R-squared values and lower RMSE values when compared to ANN. The primary reason for this performance difference lies in ANFIS’s hybrid structure, which combines the advantages of fuzzy inference systems and neural networks [47,49]. The superiority of ANFIS becomes particularly evident in areas such as modeling nonlinear data [29,47] and effectively managing uncertainties. ANFIS offers a more flexible model than ANN due to its integration of fuzzy inference systems.

As shown in Figure 15 and Figure 16, it was observed that FSACO was more effective than FSANNCO in capturing the dynamic changes in the welding arc during transitions between different current levels. In contrast, ANN models are more sensitive to model parameters such as the number of layers, number of neurons, and learning rate. As a result, their performance may decline, particularly when dealing with complex and uncertain datasets. Since ANN lacks an integrated fuzzy logic component, it may not effectively process the nuances that ANFIS can capture. Furthermore, ANFIS has the practical advantage of being easier to implement. The creation of fuzzy rules does not require extensive experience with the process [37]. In conclusion, the hybrid structure, flexibility, and ability to manage uncertainties are the key reasons why ANFIS outperforms ANN. These features make ANFIS a more suitable choice for modeling nonlinear data and analyzing dynamic systems.

4. Discussion

This study provides essential information on the use of arc light signals emitted during electric arc welding for current estimation. The results obtained are consistent with previous studies, as referenced in [1,22,47]. In this study, based on previous ANFIS based welding estimation studies [47,49], a new arc light current estimation approach is presented. Gokmen et al. [47] developed a method for power consumption estimation with sound data. In the study, the proposed method estimates current using arc light. In previous studies in the literature, power consumption estimation with sound data [47] and weld bead width and penetration depth estimation with IR thermal image data [49] were performed using ANFIS. Our findings are consistent with the studies [47,49] supporting the applicability of ANFIS in welding systems.

The study highlights the critical role of arc light in the welding process and demonstrates that a successful estimation model can be created between arc light and current. Weld quality, especially in processes requiring manual welding, is related to the welder’s experience and skill in visual perception. Therefore, visual perception has the advantage of providing a wealth of information without contacting the workpiece or the welding circuit [6]. During the welding process, welders manually adjust the welding distance speed and other parameters depending on the arc light, and these changes affect the current drawn from the electrode welding machine. In this context, our study shows that estimating current using arc light can automate the process independently of human factors and that this approach can be beneficial for monitoring the dynamics of the welding process and detecting anomalies. Additionally, it can also be used to enhance efficiency, reduce error rates, and ensure process control in welding automation, robotic welding, and artificial intelligence-based welding systems. In the study, it is observed that ANFIS models created with training and test data provided high compatibility with FCSs but showed a certain deviation and delay against sudden changes.

The strength of this study compared to previous works in the literature lies in the use of arc light for welding current estimation and the use of the ANFIS method for current estimation. However, the study also has limitations. Increasing the number of datasets at different current levels suggests that the diversity of results can be expanded. Additionally, the diversification of various filtering and artificial intelligence methods, as well as different arc welding techniques and applications, could enhance the scope and richness of the study.

5. Conclusions

In this study, the usability of arc light signals emitted during electric arc welding in the estimation of the welding current is investigated. A system was developed that allows the light sensor to move synchronously with the electrode holder for data acquisition. The arc light and current data acquired at maximum current levels of 140 A, 120 A, 100 A, and 80 A, and with two different electrode diameters, were analyzed using filtering and shifting processes. These processes ensured that the signals were denoised, data quality was improved, and estimation accuracy was increased. The ANFIS model was trained using the obtained arc light and current data, and then, current estimations were performed on the test data. The model results were enhanced through filtering, shifting, and current-limiting processes. The performance of the ANFIS models was evaluated with cross-correlation and R-squared values. Gernerally, the highest R-squared and cross-correlation values were achieved for the training and testing phases in ANFIS models, as shown in Table 2. ANN models showed lower R-squared values in both training and testing phases.

As a result of comparing the performance metrics of the ANFIS and ANN models, better estimation results were obtained for ANFIS when compared to ANN. The study shows that successful results can be obtained in current estimation by effectively using the arc light signals emitted during the welding process with artificial intelligence methods.