1. Introduction
Recently, as carbon neutrality policies have intensified globally, major countries such as France, the United Kingdom, and Norway have announced legislation to ban the sale of internal combustion engine vehicles, thereby accelerating the proliferation of Electric Vehicles (EVs)1,2). This shift has heightened technological interest in ensuring the efficiency and reliability of manufacturing processes for batteries, which are core components of EVs. An EV battery system is structured into cells, modules, and packs; the most fundamental unit, the cell, consists of an anode lead tab made of aluminum (Al) and a cathode lead tab made of copper (Cu). These cells are connected in series or parallel and joined to a copper busbar to form a module3,4).
Ultrasonic, resistance, and laser welding are employed to join these lead tabs and busbars; among these techniques, laser welding is rapidly gaining traction in industrial applications due to advantages such as its non-contact nature, high precision, and excellent compatibility with automated processes3-7). In particular, as EV demand grows and manufacturing processes become more sophisticated, the importance of laser welding in the secondary battery industry has become increasingly prominent. However, due to the significant dissimilarities in the physical properties of Al and Cu-such as high reflectivity (Al: approx. 96%, Cu: approx. 90%), different melting points (Al: 685°C, Cu: 1,036°C), and differing thermal conductivities (Al: 247 W/m·K, Cu: 398 W/m·K)-their laser welding is susceptible to various defects, including molten pool instability, porosity, hot cracking, and the formation of intermetallic compounds (IMCs)8-18).
Meanwhile, efforts are also underway to quantitatively evaluate welding quality and establish standardized criteria to ensure the reliability of the joining process and enhance its industrial applicability.
To secure the reliability of these laser welding processes and increase their industrial applicability, attempts have been made internationally to quantitatively assess welding quality and establish standardized criteria19,20). In contrast, domestically, there is a lack of certified quality standards for Al-Cu dissimilar laser welding, and many companies rely on internal standards, such as visual inspection or criteria based on tensile/shear strength per unit length (e.g., a minimum of several kgf per 1 mm)4,8). This qualitative and post- process evaluation-oriented approach to quality management makes it difficult to detect welding defects in real-time, which can lead to issues such as increased rework, higher scrap rates, and reduced productivity.
Consequently, technologies are recently being developed to analyze and predict real-time quality based on the optical and thermal signals from the plasma generated during laser welding, and these are gaining attention as core foundational technologies enabling a shift towards automated quality management systems. Sibillano experimentally demonstrated that by measuring the plasma spectrum emitted during laser welding in real-time with a spectrometer-based optical sensor, it is possible to detect penetration depth, bead geometry, and the presence of defects by analyzing the correlation between electron temperature and specific spectral lines21). Eriksson and She et al. utilized a multi-sensor system including photodiodes and IR sensors to simultaneously collect plasma emission and thermal signals generated during laser welding, experimentally proving that changes in penetration depth and bead geometry can be detected in real-time22,23). Furthermore, in the case of Al-Cu dissimilar laser welding, research in optodynamic process control has used photodiodes to measure the intensity of vapor plume emissions from aluminum during welding, enabling the real-time identification of excessive Al melting and, based on this, the classification or prediction of key quality indicators24).
While numerous studies have been conducted on the real-time analysis of external quality characteristics such as penetration depth and bead geometry during dissimilar material laser welding, research on the real-time prediction of mechanical properties, such as the shear strength of the weld, and its quantitative quality evaluation remains insufficient. Particularly for the welding of Al-Cu dissimilar busbars used in EV battery systems, attempts to quantify mechanical quality based on real-time sensor signals and link it to predictable quality standards are scarce. Therefore, a systematic approach is needed to investigate the correlation between the physical and optical signals collected during laser welding and the strength characteristics of the weld, and to develop a framework for real-time prediction and evaluation of mechanical quality based on this relationship.
Accordingly, this study proposes a quality monitoring technique capable of quantitatively predicting welding quality in real-time, based on the plasma emission and optical signals generated during the Al-Cu dissimilar dual-beam laser welding process. To this end, multi-channel optical data were collected from non-contact sensors, namely a spectrometer and a photodiode, and a training dataset was constructed by integrating this data with Fast Fourier Transform (FFT)-based energy spectra and process parameters. Subsequently, a multi-output Deep Neural Network (DNN) model was designed and trained to simultaneously predict three quality characteristics: tensile strength (regression), penetration mode (multi-class classification), and the presence of a gap (binary classification).
2. Experimental Methods
2.1 Materials and Experimental Setup
The specimens used in this study, as shown in Fig. 1, had dimensions of 45 mm × 45 mm with a thickness of 0.4 mm, and lap welding was performed using an upper plate of EN-AW 1050 Aluminum (Al) and a lower plate of C1020-HO Copper (Cu). The overlap distance was 15 mm, and each experiment was conducted with a weld line length of 35 mm. The dual-beam laser welding system used for the experiments was a single-mode dual-beam laser welder (YLS 1000/1000 SM AMB, IPG), as depicted in Fig. 2. Detailed specifications of the equipment are provided in Table 1.
Table 1
Dual beam laser welding machine spec.
| Model | YLS-1000/1000-SM-AMB | |
|---|---|---|
| Laser type | Core beam | Ring beam |
| Laser power(Max.) | 1,000 W | 1,000 W |
| Fiber cable size | 14 um | 100 um |
| Spot size | 32 um | 228 um |
| Laser delay time | 150 us | |
| Laser wavelength | 1,070 nm | |
The diagram on the right side of Fig. 2 illustrates the system used for real-time laser welding monitoring in this study, with the detailed monitoring structure shown in Fig. 3. The real-time laser welding monitoring system was composed of spectrometer and photodiode sensors. The laser welding monitoring sensor is a system that collects in real-time the optical signals spontaneously emitted from the plasma and high-temperature molten zone formed during welding; the spectrometer sensor measured wavelength-specific light intensity information by dividing the visible light spectrum into four channels, with the respective detection ranges for each channel being 382-434 nm, 484-523 nm, 572-607 nm, and 642-682 nm. The photodiode sensor consists of two channels corresponding to the laser wavelength and the near-infrared (NIR) wavelength, with respective center wavelengths of 1,064 nm (Laser PD) and 1,180 nm (NIR PD). The optical data from a total of these six channels were utilized to predict welding quality.
2.2 Data Collection Experiments and Model Design
In this study, a Full Factorial Design (FFD) was employed for the experimental design to train the DNN. FFD offers the advantage of enabling a quantitative analysis of not only the main effects but also the interaction effects among factors by conducting experiments across all level combinations of each factor; this approach is particularly effective in manufacturing processes like laser welding, where complex and nonlinear interactions among process variables occur. Particularly when constructing a DNN prediction model based on multivariate manufacturing data, such as that from welding processes, systematically designing the input variables and their combinations through the Design of Experiments (DoE) and constructing a dataset that reflects their interactions plays a crucial role in enhancing the model’s performance25). Accordingly, the primary control variables for this experiment were four in total-core beam power, ring beam power, welding speed, and gap distance-and for the first set of experiments, an FFD was designed focusing on basic combinations considering the interactions between variables, while in the second and third sets, the ranges for power and speed were expanded using FFD, with a total of over 500 experiments conducted. The dataset constructed through these experiments contributed to ensuring the diversity and reliability of DNN training. The variable levels are summarized in Table 2, and the experimental design was structured to enable a systematic analysis of how welding outcomes change with various combinations of variables, using the full factorial FFD method.
Table 2
Experimental design variable level using full factorial experiment method
2.3 DNN Model Design
2.3.1 Data Preprocessing
In this study, a total of six channels of time-series optical data were collected from the spectrometer (4 channels) and photodiode (2 channels) as shown in Fig. 3, and a series of preprocessing steps were performed to construct the input values for the weld quality prediction model based on this data. First, to remove noise, anomalous responses, and missing values from the collected raw signals, threshold filtering was applied, with data points having a signal intensity below 10 in each sensor channel treated as outliers and removed. After filtering, the remaining time-series data were converted into statistical summary values, and the mean and standard deviation were calculated for each channel. Subsequently, to extract features in the frequency domain, a FFT was performed, and from its results, energy spectrum values were derived. In addition to these sensor-based derived variables, a total of four process condition variables-ring beam power, core beam power, welding speed, and the presence of a gap-were integrated to form a final set of 22 input variables. Among these, 18 are sensor-based variables, specifically comprising the mean and standard deviation for the 4 spectrometer channels (8 variables), the mean and standard deviation for the 2 photodiode channels (4 variables), and the FFT-based frequency characteristics for all 6 channels (6 variables). The preprocessed dataset was utilized as the input data for model training and was subsequently applied to a multi-output DNN model that includes both classification and regression tasks.
2.3.2 Input Data Design for Training
It is known that the number of training samples should be at least ten times the number of input variables26), and in this study, a total of 510 experimental data points were secured for the 22 input variables. To improve the generalization performance of the deep learning model, a data augmentation technique was applied by injecting Gaussian noise with a standard deviation in the range of 0.01-0.03 into each sample to reflect realistic sensor noise while preserving the original data distribution. Through this process, a total of 5,100 training samples were constructed, representing a tenfold expansion of the original dataset. Although this does not meet the ideal criterion proposed by Alwosheel of having more than 50 times the number of parameters27), it is considered a sufficient data scale for evaluating the weld quality classification and tensile strength prediction model using diode and photodiode sensors in this study..
2.3.3 DNN Architecture Design
For the multi-output deep neural network architecture designed in this study, the optimal structure was derived by employing the Random Search technique within a defined parameter search space for the number of hidden layers (2-4), the number of nodes per layer (128-512), and the dropout ratio (0.1-0.3). This model follows a Multitask Learning architecture that shares internal hidden layers based on a single input layer and can simultaneously perform multi-output predictions for three quality variables: weld quality status (complete, partial, no joint), tensile strength (kgf), and the presence of a gap (present, absent). This model is composed of a structure with common hidden layers and branched output layers to effectively share interaction information among input features and enable the learning of relationships between the outputs.
Following the input layer, two hidden layers were sequentially placed, and to prevent overfitting, Batch Normalization, Dropout (ratio 0.1-0.3), and L2 Regularization (λ = 1×10-4-1×10-3) were applied in conjunction with the ReLU activation function in each layer. Subsequently, the three branches derived from the common hidden layers were each configured with an independent structure tailored to their respective prediction tasks.
The weld quality classification branch is composed of one hidden layer (128 nodes) and an output layer with a Softmax function (for 3 classes: complete, partial, no joint), and was trained using the Sparse Categorical Cross-Entropy loss function.
The tensile strength prediction branch, designed to effectively model a more complex regression function, includes two hidden layers with 256 and 128 nodes, respectively, and an output layer with a linear activation function, using Mean Squared Error as its loss function. The gap presence detection branch consists of a hidden layer with 128 nodes and an output layer with a Sigmoid function, and was trained using the Binary Cross-Entropy loss function.
The loss values for each output were adjusted with a weight ratio of 1:1.5:1 for weld quality, tensile strength, and gap, respectively, and the model was optimized using the total weighted loss, a design intended to balance the prediction performance across the individual outputs. Optimization was performed using the Adam Optimizer (learning rate 5x10-4), and was complemented by Early Stopping (patience=30) and a learning rate reduction condition (ReduceLROnPlateau, patience=20, factor= 0.3) to prevent overfitting.
Model training was conducted using 5,100 augmented samples, randomly split into training and testing data at a 7:3 ratio, and was set to terminate only when a weld quality classification accuracy of over 92% and a tensile strength regression coefficient of determination (R2) of over 0.82 were simultaneously achieved. If these conditions were not met, automatic retraining was repeatedly performed under the same conditions to ensure optimal performance was secured. The primary hyperparameters for training set in this model were epochs=3000 and batch_size=16.
As such, the multi-output neural network with the architecture shown in Fig. 4 achieved simultaneous improvements in prediction efficiency and generalization performance compared to conventional single-output prediction structures, by enabling parameter reuse and mutual learning synergy.
3. Results
3.1 Evaluation of the Trained Model
3.1.1 Evaluation of the Trained Model
In this study, to quantitatively evaluate the performance of the trained prediction model, the coefficient of determination (R2) was utilized for the regression model, while different metrics were applied for the classification models depending on the problem type.
For the prediction of weld quality status, a multi-class classification task, the confusion matrix, F1 score, and accuracy were used as the main performance indicators, whereas for the prediction of gap presence, a binary classification task, precision and recall were comprehensively reviewed.
Among these, the F1 score is defined as the harmonic mean of the classification model’s precision and recall, and it is known as a suitable metric for evaluating the balanced predictive performance of a model, even in situations of class imbalance. Each evaluation metric is defined as shown in Equations (1)-(5)28-30).
Here, True Positive (TP) refers to the number of actual positive samples correctly predicted as positive, while False Positive (FP) signifies cases where an actual negative sample is incorrectly predicted as positive. False Negative (FN) represents cases where an actual positive sample is mistakenly classified as negative, and True Negative (TN) refers to cases where an actual negative sample is correctly predicted as negative.
The performance of the trained prediction model was evaluated to be at a high level in both the classification and regression tasks. For the multi-class classification of weld quality, the overall accuracy was 92.3% and the average F1 score was 0.89, indicating that balanced prediction among the classes is possible. In the case of gap presence prediction, the Precision was measured at 0.91 and the Recall at 0.87, demonstrating that both FP and FN are being stably suppressed.
Meanwhile, in the tensile strength prediction by the regression model, the coefficient of determination (R2) was derived at an average level of 0.82, which signifies that the trained model possesses significant explanatory power in quantitatively predicting material properties from the complex patterns of optical signals. This result is interpreted as experimental proof that the designed DNN model effectively learned from high-dimensional sensor-based signals and is practically applicable for predicting quality in actual welding processes.
3.1.2 Weld Quality Classification Results
The weld quality status classification was trained on three classes-complete penetration, partial penetration, and no joint-and the test results showed an overall accuracy of 93.5% and an F1 score of 93.32%. The confusion matrix results are presented in Fig. 5; for complete penetration, a high discrimination performance was maintained, and for no joint, the prediction precision was the highest, confirming the model’s high sensitivity to this type of defect. In contrast, the partial penetration class exhibited a relatively lower recall, which is presumed to be due to either an ambiguous definition of the class or an insufficient amount of input data for training.
3.1.3 Gap Presence Detection Results
The presence or absence of a gap occurring during the welding process was determined in a binary classification format, and the prediction results for the test data are as follows. The gap classification results showed high performance, with an accuracy of 97.4%, an F1 score of 97.8%, a precision of 98.9%, and a recall of 96.8%; the confusion matrix results are presented in Fig. 6.
The very high precision of 98.9% indicates that there were few cases incorrectly identified as False Positives (FP) despite the absence of a gap. Furthermore, the high recall of 96.77% means that the model successfully detected the majority of samples where a gap was actually present. Since both precision and recall are high, the F1 score was also high at 97.83%, which implies that there was no significant data imbalance or learning instability between classes.
According to the confusion matrix analysis, misclassification cases in the gap presence classification were very few; in particular, when classifying the “absent” gap state, the recall was over 98%, indicating an excellent ability to accurately classify the normal state.
Fig. 7 visualizes the predicted probability distribution of the gap classification results, showing that most samples are distributed in extreme probability intervals close to 0 or 1. This suggests that the model performs clear binary classification with high confidence for most samples, rarely making predictions in uncertain intermediate probability ranges such as 0.4-0.6 regarding the presence of a gap. In particular, samples without a gap are concentrated in the 0.0-0.1 interval, while samples with a gap are concentrated in the 0.9-1.0 interval, confirming that the separability of the probability distributions between classes is excellent. This indicates that the model’s classification boundary is distinct and the confidence in its prediction results is very high.
3.1.4 Tensile Strength Prediction Results
In this study, to evaluate the tensile strength prediction performance, predictions were performed using the DNN model on 100 tensile strength data points randomly sampled from the total of 510 acquired through experiments. The comparison results between the predicted and measured values are shown in Fig. 8, and the model’s coefficient of determination (R2) is 0.8367. This level of performance, capable of explaining approximately 83.7% of the total variance, is evaluated as having secured high reliability as a sensor-based weld quality prediction model. Generally, an R2 value of 0.8 or higher is considered to indicate an excellent predictive power for a regression model, and particularly in a real welding process environment with non-linearity and disturbances, the model is judged to be reliable if a coefficient of determination of 0.75 or higher is secured31). However, in Fig. 8, an increase in error was observed in the segment after Sample No. 60; this is believed to have occurred because the variability of tensile strength was relatively large in this segment due to changes in welding process conditions, and insufficient experimental data under similar conditions were secured during the model training process. Therefore, it is thought that the prediction performance in this segment can be improved in future research by conducting additional experiments to increase the data density.
Consequently, the prediction model of this study is considered to have effectively learned the FFT-based energy spectra and process parameters (beam power, travel speed, gap presence, etc.) and to have stably modeled the nonlinear relationship with tensile strength.
4. Conclusion
This study proposed a multi-output Deep Neural Network (DNN) model capable of quantitatively predicting weld quality in real-time based on optical signals generated during Al-Cu dissimilar dual-beam laser welding, a key process in electric vehicle battery systems, and experimentally validated its effectiveness. In particular, it identified the quantitative correlation between physical signals collected via non-contact sensors and mechanical quality characteristics, and implemented this in a structure capable of real-time prediction.
1) A total of six channels of time-series data were collected from the plasma and thermal emissions generated during welding using a spectrometer (4 channels) and a photodiode (2 channels), and the signal from each channel was quantified into features such as mean, standard deviation, and FFT-based energy spectrum. These were integrated with process variables to construct a total of 22 input features.
2) The proposed model was designed with a multi-output DNN architecture to simultaneously perform three prediction tasks-weld quality status classification (3 classes), tensile strength prediction (regression), and gap presence detection (binary classification)-and stable training performance was secured by applying data augmentation and overfitting prevention techniques.
3) The trained model demonstrated favorable performance in both classification and regression predictions, achieving a coefficient of determination (R2) of 0.8367 for tensile strength prediction. Furthermore, for weld quality status classification (3 classes), it achieved an average accuracy of 92.1% and an F1 score of 0.913, while for gap presence detection (binary classification), it was evaluated with an accuracy of 96.8% and an F1 score of 0.966, securing excellent predictive performance for each quality item.
This study experimentally demonstrated that time-series signals collected from optical-based sensors have a statistically significant correlation with mechanical quality characteristics and implemented this into a system capable of real-time quantification and prediction using an artificial intelligence model. If data expansion to include more diverse process conditions and materials, along with model enhancement, is pursued in the future, the quality prediction technique proposed in this study is expected to be applicable as a foundational technology for real-time quality management and process control automation in electric vehicle battery manufacturing sites.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
