1. Introduction
Due to recent environmental regulations, the electric vehicle industry has experienced rapid growth, resulting in a surge in the production of secondary batteries. In this process, both similar and dissimilar welding of aluminum and copper has become a key step in battery manufacturing. While ultrasonic welding was primarily used in the past, laser welding is now being actively adopted due to its advantages such as high speed and precision1,2). Laser welding involves focusing high energy onto a localized area, which results in complex thermo-fluid phenomena such as keyhole and plasma formation, as well as rapid solidification of the material. These phenomena can cause various defects, such as porosity, cracking, and spatter, either within the weld or on the surface-especially when welding difficult materials like aluminum and copper1,2). In particular, poor weld quality in tab-to-tab or tab-to-busbar joints in secondary batteries can lead to mechanical and electrical failures, posing risks such as thermal runaway3).
As a result, a systematic management system is required to ensure high-quality welds and defect detection during manufacturing. In secondary battery production, it is essential to implement a monitoring system capable of real-time defect detection and early identification of quality anomalies during laser welding processes. In response, various sensor-based quality evaluation technologies suitable for laser welding process monitoring are being actively studied4,5). Among these, optical-based monitoring methods using cameras, photodiodes, and spectrometers are widely employed due to their non-contact and high-precision measurement capabilities6).
High-speed cameras can visually capture anomalies during welding by recording phenomena such as plasma plumes, spatter, and changes in the molten pool at ultra-high speeds. In contrast, visible-light (VIS) photodiodes offer the advantage of quantitatively tracking the intensity and variation of plasma by measuring the amount of visible light emitted.
Previous monitoring studies primarily relied on single-sensor approaches using photodiodes, thermal sensors, or cameras to measure data and evaluate quality7,8). However, single sensors are limited in their ability to capture complex phenomena since they only measure one physical variable. Recently, research has increasingly focused on multi-sensor monitoring systems, which enable multifaceted analysis by integrating cross-sensor data and applying algorithms such as machine learning5). Nevertheless, there has been a relative lack of research aimed at understanding physical phenomena or identifying defect causes based on multi-sensor data. Moreover, although some studies have explored the correlation between multi-sensor data in laser welding and its use in anomaly detection or quality assessment9,10), there are few cases where algorithms have been designed or optimized for real-time application.
Against this backdrop, the present study develops a multi-sensor monitoring system using a high-speed camera and a VIS photodiode to evaluate keyhole and molten pool behavior, as well as defect characteristics, in aluminum-copper dissimilar laser welding. The entire welding process was divided into nine segments, and brightness and area metrics were extracted from the high-speed video. These are then compared with irradiance data optical intensity data acquired from the VIS sensor. Pearson and Spearman correlation analyses were conducted between the image-based metrics and VIS sensor data to eliminate redundant variables with similar trends, thereby enhancing interpretability and confirming the correlation between key metrics.
2. Laser Welding Experiment and Methods
2.1 Materials and Equipment for Laser Welding
To observe the phenomena and signal changes associated with the welding characteristics of aluminum-copper dissimilar welding for secondary batteries, a series of experiments was conducted. For the laser welding tests, Al 1050 was used as the top plate and Cu 1100 as the bottom plate. The chemical properties of each plate are shown in Table 1, the specimen geometry in Fig. 1, and the configuration of the experimental setup in Fig. 2. The laser used in the experiment was a 4 kW multi-mode fiber laser (Trudisk 4000, Trumpf). The welding process was recorded using a high-speed camera (VEO410L, Phantom) at 80,000 frames per second (FPS), positioned at a 45° angle and a distance of 350 mm to capture both the keyhole and plasma plume. The plasma ejected from the keyhole was also measured using a photodiode sensor. The sensor was positioned at a 60° angle and a distance of 150 mm, with a sampling rate of 10 kS/s, to capture optical intensity data from both the keyhole and plasma plume. The welding was performed by moving the bed so that the welding point remained fixed in 3D space, allowing both the camera and photodiode to capture the entire welding process.
2.2 Welding Conditions and Data Extraction Method
The experiment was carried out in a lap joint configuration, with a 0.4 mm gap introduced between the top and bottom plates to intentionally create defective conditions. The welding parameters are listed in Table 2. To ensure data reliability, each experiment was repeated three times.
Table 2
Laser welding condition
| Parameters | Conditions | |
|---|---|---|
| Laser power | Core | 0.85 kW |
| Ring | 0.85 kW | |
| Welding speed | 80 mm/s | |
| Beam wavelength | 1030 nm | |
| Gap | 0.4 mm | |
For image data analysis, the following metrics were extracted from the recorded images: overall image brightness, plasma brightness, molten pool area, and keyhole area. The data extraction process is illustrated in the flow chart in Fig. 3.
During the processing of high-speed video data, resizing degraded image quality. To enhance visibility and analytical precision, an upscaling technique was applied, increasing the number of pixels by a factor of 10 compared to the original resolution. New pixels were generated using various upscaling methods, including nearest-neighbor interpolation, bilinear interpolation, and bicubic (third-order convolution-based) interpolation. These methods were considered as representative approaches for the upscaling process11). The results of applying these methods to the image data are shown in Fig. 4. Nearest-neighbor interpolation was the fastest, while bicubic interpolation was the slowest. As the complexity of the interpolation method increased, the processing time increased accordingly, with up to 33–36 seconds of additional computation time required. In terms of image quality, bilinear and bicubic interpolation produced similar results, whereas nearest-neighbor interpolation introduced visible stair-step artifacts. Therefore, bilinear interpolation was selected for its favorable balance between processing speed and image quality, making it suitable for real-time application.
Fig. 4
Comparison of image quality according to different upscaling methods, (a) Nearest-neighbor, (b) Bilinear, (c) Bicubic
Next, the following features were extracted: image gray level (IG), plasma gray level (PG), molten pool area (MPA), and keyhole area (KA). The overall image brightness (IG) was calculated as the average gray level of pixels within a predefined region of interest (IG ROI), while plasma brightness (PG) was calculated as the average intensity of the blue channel within the PG ROI. To measure the 2D area of the molten pool and keyhole observed from the top view, separate ROIs were defined.
The area of each region was estimated by detecting edges, constructing polygons based on those edges, and counting the number of pixels within each polygon. To improve edge detection accuracy, the ROI for 2D area extraction was divided into an 8×8 grid of uniform blocks, and edge coordinates were extracted within each block using the designed algorithm. A polygon was then constructed using the extracted coordinates and the monotone chain algorithm, and the sum of the pixels enclosed by the polygon was taken as the area of the molten pool or keyhole. Fig. 5 shows the IG and PG ROIs and the resulting polygon shapes extracted through edge detection. However, since the keyhole and plasma are three-dimensional physical phenomena, 2D area-based analysis alone is limited in capturing the full complexity of the process. In future work, volumetric analysis using multi-angle video data could enable quantitative analysis of spatial behaviors such as plasma ejection direction and volumetric oscillation, allowing for more precise interpretation of the underlying physical mechanisms.
3. Analysis of Welding Section Characteristics Based on Image and Photodiode Sensor Data
3.1 Definition of Welding Sections and Image Pre- processing
To analyze specific characteristics in different parts of the welding process captured by the high-speed camera, the welding process was divided into nine segments, from the start of welding to the reformation of the keyhole, as shown in Fig. 6. Fig. 6 (a) and (b) correspond to the initial phase of welding, during which the laser power is still ramping up to the target experimental condition. Because the heat input and energy density are lower than the threshold required for keyhole formation, the welding proceeds in conduction mode. Fig. 6 (c) represents the transition from conduction mode to keyhole mode as the laser power reaches levels close to the experimental condition. Fig. 6 (d) and (e) are where stable keyhole-mode welding occurs under the full experimental laser power. In segment (d), the keyhole remains stable throughout the weld, while in segment (e), the keyhole begins to destabilize due to the presence of a gap between the plates. Fig. 6 (f) marks the transition where the unstable keyhole collapses. In segment (g), after the keyhole collapses, the plasma plume ceases, and melting primarily occurs on the surface of the top plate. Fig. 6 (h)and (i) capture the reformation of the keyhole, with the plasma plume ejected at a diagonal angle opposite the welding direction.
Fig. 6
Comparison between high-speed images and specimen for segment selection in the welding process, (a) Welding start, (b) Conduction mode welding, (c) Conduction-keyhole transition, (d) Keyhole mode welding, (e) Keyhole instability, (f) Keyhole collapse transition, (g) Keyhole collapse, (h) Keyhole collapse-reformation transition, (i) Keyhole reformation
For each segment, time axes were synchronized, and the number of data points for analysis was standardized: 360 images from the high-speed camera and 45 optical intensity data from the VIS photodiode sensor per segment. To analyze the trends and correlations of image-derived features (IG, PG, MPA, KA), it was necessary to remove noise. Several filtering techniques were considered, including Gaussian, median, and bilateral filters. Among these, the Gaussian filter showed the least discontinuous patterns and localized distor- tions. Since the main objective of this study was to compare general trends across welding modes and to perform correlation analysis, it was more important to preserve overall signal patterns than to capture fine signal fluctuations. Therefore, the Gaussian filter was selected. The standard deviation σ is a key parameter in the Gaussian filter, controlling the smoothing intensity. A larger σ generally improves noise removal but can also remove useful signals. To select an optimal σ, the noise reduction ratio (NRR) was calculated for each σ value, as shown in Table 3. Due to the structural properties of the Gaussian filter, increasing σ results in smoother data and reduced NRR. However, such a decrease in the NRR does not indicate that only noise has been removed; it may also imply the loss of useful signals. Thus, the absolute value of NRR alone is insufficient for assessing useful signal loss or selecting the optimal σ. An additional criterion is therefore necessary to guide parameter selection. Accordingly, when the decrease in NRR drops below 20% compared to the previous σ value for all metrics, this was considered saturation in noise reduction, and further increases in σ were deemed unnecessary. The analysis showed that σ = 3 met these conditions and was thus adopted. Finally, since the number of data points from the image and sensor data did not match within each segment, linear interpolation was applied to match the data counts before performing analysis.
3.2 Conduction Mode Sections
Fig. 7 (a) shows that, after laser irradiation, the molten pool began to form via conduction. The molten pool was small and highly stable in shape. Fig. 7 (b) shows that the molten pool area increased, and its boundary became more defined. A stable and well-defined plasma plume was observed ejecting upward. As a result of these phenomena, all metrics from both the image and photodiode sensor data increased by approximately 10-20% compared to Fig. 7(a). Additionally, since no keyhole was formed in this segment, the keyhole area metric remained close to zero.
Fig. 7 (c) shows that, as the conduction mode molten pool approached the transition point, it became increasingly unstable, and all data metrics dropped by approximately 10% compared to stable conditions. The transition from conduction mode to keyhole mode occurred within an interval of approximately 87.5 μs, and both the molten pool and keyhole area metrics began to increase at the moment of transition. After the transition, the metrics across all data sources exhibited high variance, particularly the plasma brightness, which showed significant fluctuation. This instability is attributed to the difficulty in maintaining a stable keyhole during the transition phase, leading to the observed data variations.
3.3 Keyhole Mode Sections
Fig. 8 (a) and (b) show that data fluctuations were significantly greater in the keyhole mode compared to the conduction mode. In keyhole mode, keyhole oscillations are induced by vapor pressure within the keyhole and thermal convection driven by the Marangoni effect acting along the keyhole walls12,13). As a result, unlike conduction mode welding-where the plasma plume is stably ejected-the direction of plasma plume ejection in keyhole mode fluctuates in response to keyhole oscillation, causing increased variability in the measured data.
Fig. 8 (a) shows that, unlike the unstable transition phase immediately after keyhole formation, the molten pool area did not differ significantly from that observed in the conduction mode. However, brightness metrics-such as image brightness, plasma brightness, and photodiode sensor data-showed relatively large fluctuations. Although the amount of metal melted in keyhole mode is greater than in conduction mode (which should increase light emission from the keyhole and plasma plume), the brightness metrics were lower. This is likely due to the dispersion of the plasma plume caused by keyhole oscillation, leading to greater variation in brightness measurements.
Fig. 8 (b) shows that the presence of a gap caused instability in the keyhole, resulting in significant fluctuations in both brightness and area metrics. The keyhole temporarily collapsed and was re-formed over a 15 μs interval. During this collapse, all metrics-except for the photodiode sensor signal-dropped to nearly zero. Although the photodiode measured values in the 2-3 V range during the collapse period, the lower sampling rate of the sensor compared to the image data and its sensitivity to the Gaussian filter led to a slight drop in voltage, as shown in the graph.
3.4 Keyhole Collapse Section
Fig. 9 (a) shows that the keyhole was extremely unstable just before collapse, and among all keyhole mode segments, this segment exhibited the highest data fluctuations. The keyhole collapse transition occurred within an interval of approximately 125 μs, during which all metrics rapidly decreased. Immediately after the transition, the plasma plume disappeared, and all data values dropped close to zero.
Fig. 9 (b) shows that, following the keyhole collapse, the plasma plume was no longer observed, and all data except for image brightness converged toward zero. Data analysis showed that the top plate melted more extensively around the laser irradiation zone. Light emission at the melt boundary caused a slight increase in image brightness. Surface bead morphology analysis of the specimen revealed that the top plate had melted and was displaced outward around the weld line. This is attributed to the keyhole collapse, which prevented sufficient melting of the bottom plate and limited heat conduction into it. As a result, residual thermal energy diffused through the gap and into the top plate, leading to the observed melt pattern on the surface.
3.5 Keyhole Reformation Section
Fig. 10 (a) shows that a semi-circular keyhole formed in the forward welding direction, and a plasma plume was observed stably ejecting diagonally in the reverse direction. This suggests that, after the keyhole collapsed, a broad molten pool formed around the weld line on the top plate. During the reformation process, thermal imbalance led to the internal formation of the keyhole in a diagonal direction and caused the plasma to eject accordingly. Due to the stable plasma plume, data fluctuations were minimal, and the values were similar to those observed at the beginning of the welding process (Fig. 6(a)). However, the brightness metrics and photodiode sensor values in Fig. 10(a) were significantly lower.
Fig. 10 (b) shows that during the complete reformation of the keyhole, it exhibited pronounced oscillations due to instability. As the reformation progressed, the volume of molten metal increased, which led to a rise in plasma plume brightness. However, the plume ejected diagonally toward the rear became increasingly unstable. As a result, fluctuations in all data metrics increased. Notably, the plasma brightness was measured to be particularly high, as the plasma plume was concentrated around the diagonal rearward direction.
4. Correlation Analysis Between Sensor Data
In a multi-sensor-based monitoring system, the addition of more sensors leads to an increase in data inputs for the analytical model, potentially causing various issues. First, physical interactions between sensors may occur, meaning that a malfunction in one sensor could cascade and affect others, thereby reducing the reliability of the data from each sensor14). Additionally, metrics with similar trends but low analytical significance may emerge, making it difficult to identify key metrics. Among various techniques to address these issues-such as calibration, synchronization, and normalization-correlation analysis is essential for performing multi-sensor monitoring. By analyzing correlation coefficients under both normal and faulty conditions, correlation analysis enables early detection of defects or hardware anomalies14). It can also be used to reduce redundant data with similar trends, thus simplifying the analytical model. Complementary data can be weighted differently and used as key analytical features.
In this section, the average values of each data-brightness and area metrics (IG, PG, MPA, KA) extracted from high-speed camera images and the VIS photodiode sensor values-were used to calculate Pearson and Spearman correlation coefficients. Based on the interpretation ranges proposed by Schober et al.15), the degree of correlation was classified into five levels from “Negligible” to “Very Strong”, as shown in Table 4, and the correlation coefficients were evaluated accordingly.
Table 4
Interpretation of correlation coefficients based on magnitude
| Coefficient magnitude | Interpretation (Correlation) |
|---|---|
| 0.00 – 0.10 | Negligible |
| 0.10 – 0.39 | Weak |
| 0.40 – 0.69 | Moderate |
| 0.70 – 0.89 | Strong |
| 0.90 – 1.00 | Very Strong |
Table 5 presents the results of Pearson and Spearman correlation analyses between all image-based metrics and VIS sensor data. Except for the combination of keyhole area and the photodiode sensor, all other combinations yielded p-values below 0.012,
Table 5
Results of correlation analysis between image-based metrics and VIS data
indicating statistically significant correlations at a confidence level of at least 98.8%. The VIS data showed meaningful correlations with all image-based metrics In particular, the image and plasma brightness exhibited a “Very Strong” linear correlation with VIS data, indicating very strong linearity. Plasma brightness had the highest correlation coefficient, approaching 1.00, suggesting the highest degree of spectral match with the photodiode sensor. Image brightness also showed a high correlation with the VIS data, though lower than that of plasma brightness. Since image brightness and plasma brightness exhibit similar trends, image brightness can be regarded as redundant. Therefore, monitoring brightness changes during the process using either plasma brightness alone or the VIS sensor alone would be feasible. Given the very strong correlation between the VIS data and plasma brightness, it is likely that any defect in either sensor could be detected in real time by analyzing residuals in their values. That is, since both metrics maintain a strong correlation under both normal and abnormal conditions, an algorithm can be designed to detect hardware faults when the difference between the two exceeds a certain threshold. This would allow for early detection of cascading sensor failures caused by physical interactions between sensors.
Regarding the molten pool area, a “Very Strong” linear correlation with the VIS sensor was observed, though the trend correlation was slightly lower, falling closer to “Strong”. For the keyhole area, since measurements are only taken from the conduction-to-keyhole transition phase onward, the analysis was conducted using data from seven sections. Results showed “Strong” linearity and “Moderate” trend correlation, similar to the molten pool area. As shown in Fig. 11(c), this discrepancy is attributed to a partial reversal in rank ordering during segments with high VIS signal intensity. For example, in conduction mode-even when the molten pool is small and metal melting is minimal-the plasma plume was concentrated and emitted vertically, resulting in high VIS sensor readings (Fig. 7(b)). In contrast, during keyhole mode, although both the molten volume and pool area increased, plasma was dispersed in multiple directions due to keyhole oscillation, reducing the amount of light captured by the sensor (Fig. 8 (a)). As shown in Table 6, during the transition from conduction to keyhole mode, the molten pool area increased by approximately 20%, while the VIS voltage dropped by about 10%. This reversal in rank order between VIS and image-based metrics explains the observed correlation patterns. Such characteristics could be useful for detecting welding mode transitions in real time. In conduction mode, the VIS/MPA ratio remains high and stable, but in keyhole mode, the ratio decreases and its variability increases. Monitoring the VIS/MPA trendline could therefore help not only identify mode changes but also detect abnormal keyhole behavior caused by defects such as incorrect laser focus. Since all area metrics show similar trends, molten pool area, which has the highest correlation, would be the most effective indicator for monitoring.
5. Conclusion
In this study, the behavior of the keyhole, molten pool, and plasma during specific segments of the laser welding process was analyzed using a high-speed camera and VIS photodiode sensor. The correlation between the sensor data was also examined. The main findings are summarized as follows:
1) In the conduction mode segment, the plasma plume was observed to be concentrated and emitted vertically, with low variation across all metrics. Brightness-related data values from each sensor were significantly high. However, upon transitioning to keyhole mode, the keyhole became highly unstable, resulting in considerable dispersion of the plasma plume. Consequently, all metric values dropped by approximately 10%, with plasma brightness showing the highest level of fluctuation.
2) In the keyhole mode segment, keyhole oscillation caused by internal vapor pressure and the Marangoni effect led to unstable plasma emission directions, significantly increasing the variation across all metrics compared to the conduction mode. Initially, in the early stage of keyhole reformation following collapse caused by a gap, both keyhole and data fluctuations remained stable. However, as the keyhole reformed further, the backward diagonal plasma emission intensified, causing all metrics to surge and exhibit greater variability.
3) Based on the average correlation analysis across segments, the VIS-PG correlation coefficient was nearly 1.00 (“Very Strong”), indicating strong linearity and trend alignment. This confirms that either the photodiode sensor or plasma brightness metric alone is sufficient for effective monitoring. The VIS-MPA correlation showed a “Very Strong-Strong” level; however, due to changes in keyhole area and plasma emission direction between conduction and keyhole modes, the trend correlation decreased. It was confirmed that the VIS/MPA trend could serve as a sensitive metric for detecting welding mode transitions.
These results demonstrate that the behavior of the keyhole and plasma under both normal and defective conditions in laser welding can be effectively analyzed using a high-speed camera and VIS photodiode sensor. Additionally, the correlation analysis offers potential for detecting hardware faults between sensors and identifying keyhole shape changes caused by welding mode shifts or defects.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
