Quality Classification of Multi-Layer Copper Foil Stacks Welds for Ultrasonic Welding Using CNN and RNN
Article information
Abstract
Abstract
This study aims to develop a model for classifying weld quality using data acquired in real-time during the ultrasonic welding process. The data utilized includes LVDT (DP-10, DAQ-Express) and power signal data. The model categorizes welds into three classes (insufficient, sufficient, and excessive) based on the total input energy of the ultrasonic welding process. To classify the quality of ultrasonic welds, Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) are employed. The focus of this research is on utilizing CNN and RNN models to classify weld quality based on signal data acquired from ultrasonic welding, aiming to enhance the reliability of secondary battery joints.
1. Introduction
Due to the recent implementation of policies aimed at regulating greenhouse gas emissions from internal combustion engines, the demand for electric vehicles (EVs) and hybrid vehicles has surged rapidly1-4). The energy density of lithium-ion batteries, which are currently the standard, is directly related to battery performance, and improving battery performance translates into longer driving ranges for electric vehicles. Therefore, enhancing the performance of the battery, which is the core driving system of electric vehicles, is the ultimate goal5). The batteries used in electric vehicles are composed of cells, modules, and packages with electrical connections at each level. The reliability of these battery connections is crucial as it can impact the risk of fire and explosion. Processes such as resistance spot welding, laser welding, and ultrasonic welding are applied to the battery joints, and in this study, ultrasonic welding was employed to join the cells6-8). Ultrasonic welding offers advantages such as being effective in controlling the micro-gaps between stacked foils due to its pressurization process and causing minimal deformation due to heat input as a solid-state welding method.
Traditional quality control methods rely on destructive and non-destructive testing after the process. To address this limitation, there is a need for a system that can gather data in real-time during the process and classify the quality of the welds. This study aims to establish a system that classifies weld quality in real-time by using displacement (Linear Variable Differential Transformer, LVDT) and power signal data obtained during ultrasonic welding. The welds were classified as insufficiently welded, sufficiently welded, or excessively welded according to the ultrasonic welding process parameters and the total input energy. Based on the acquired data, a study was conducted to classify the quality of ultrasonic welds by simultaneously applying Convolutional Neural Networks (CNN) and Recurrent Neural Network (RNN) models.
The purpose of this study is to accurately classify weld quality by effectively extracting local patterns within the time-series data through CNNs and analyzing temporal features through RNNs. This approach allows for more precise interpretation of the complex signal data generated during the ultrasonic welding process and is expected to contribute to enhancing the efficiency of the ultrasonic welding process and improving the reliability of secondary batteries.
2. Research method
In this section, we aim to explain the evaluation of ultrasonic weld characteristics in multi-layer copper foil stacks based on process parameters, the signal processing of displacement and power data, and the classification of ultrasonic welds using artificial neural networks, specifically CNN and LSTM Networks, along with an accuracy assessment based on metrics.
2.1 Experimental setup
The ultrasonic welding process is conducted as shown in Fig. 1, where joining occurs due to localized plastic deformation caused by frictional heat generated from the vibrating pressure. An ’EMC20’ ultrasonic welder from Herrmann was used. The chemical properties are presented in Table 1. Variable process parameters included applied clamping pressure (MPa), amplitude (μm), and welding time (s). Additionally, welds were classified into non-welded (Insufficient, 0), properly welded (Sufficient, 1), and overheated welds (Excessive, 2) based on the total input energy (J), as outlined in Table 2. The total input energy was calculated as power (W) x welding time (fs), and since the welder’s power could not be adjusted independently, weld characteristics were examined based on its proportionality to welding time9). The materials used were 60 layers of oxygen-free copper foil (8 μm), with dimensions of 60 mm in length and 45 mm in width.
Schematic of the ultrasonic welding process
The chemical composition in Cu-OF (wt%)
Classification of ultrasonic welds by process parameters
The evaluation of ultrasonic weld appearance focuses on the shape of the indentation, uniformity, and the presence of cracks. Cross-sectional evaluation was based on the depth of the indentation, measured by the gap between the peaks and valleys formed by the horn and anvil. The T-Peel test was conducted at a speed of 10 mm/min using a 30/30-sheet configuration, and the fracture modes were examined according to different process parameters.
2.2 Acquisition of ultrasonic welding signal data (LVDT, Power)
The DP-10 (LVDT sensor) mounted on the ultrasonic welder was connected to DAQ-Express, and LVDT data (analog signal, V) were acquired at a sampling rate of 1.63 kHz. Power data were acquired using the power controller installed inside the ultrasonic welder at a sampling rate of approximately 1.00 kHz. A schematic of this process is shown in Fig. 2. To verify whether the acquired LVDT and power signals correspond to the ultrasonic welding phenomenon, the data were plotted as graphs, as shown in Fig. 3. It was observed that the LVDT values changed as the pressing process began, and the intervals of clamping, vibration, holding, and unloading were identified. The power approached the welder’s set value of 2,000 W as welding progressed, and it was confirmed that the power dropped to 0 W at the end of the welding process10).
Schematic of LVDT and power data acqui-sition connected to an ultrasonic welder
Signal graphs of (a) LVDT and (b) power based on welding time
2.3 Classification model construction and evaluation methods
2.3.1 Characteristics of the model applying CNN and LSTM
The 1D Convolutional Layer (Conv1D) applied in this study has a structure that processes one-dimensional data in one direction, as shown in Fig. 4, and is mainly suitable for data in the time domain. The 2D Convolutional Layer (Conv2D) has a structure suitable for processing two-dimensional data (images).
The process of extracting features using Conv1D
The model incorporating CNN and LSTM networks is characterized by its ability to simultaneously learn local characteristics and long-term dependencies, allowing it to effectively handle complex data structures and achieve high prediction performance11). Local features refer to significant patterns or structures within specific locations or segments of the data, while long-term dependencies describe the relationships or patterns between different time points in time-series data, including the relationship between the early and late parts of the input data.
Traditional Recurrent Neural Networks (RNNs) are used to learn temporal dependencies in time-series data; however, they face limitations in capturing long-term dependencies due to issues such as vanishing and exploding gradients when learning from long sequences. To address this, LSTM layers were applied to retain only important information while discarding unnecessary data. Therefore, a structure was used where CNNs extract local features from the input data and pass them to the LSTM, which performs modeling that reflects temporal progression. This combination enables effective learning of changes and patterns over time, going beyond simple feature extraction.
To classify ultrasonic welds into Class 0, 1, and 2, a model connecting CNN and LSTM networks is presented in Fig. 5. The applied model has a parallel structure, allowing different input data to be learned independently, thereby minimizing bias towards specific data. This enhances the model’s ability to capture diverse patterns.
Structure of the classification model using CNN and LSTM
2.3.2 Model accuracy evaluation
The model applied in this study is a classification model, and we try to evaluate its accuracy by comparing the difference between the predicted result and the actual class for the input data based on metrics (Accuracy, Precision, Recall, F1-score) and confusion matrix12,13).
Table 3 summarizes the True Positives (TP), False Positives (FP), False Negatives (FN), and True Negatives (TN). True Positives refer to instances where the model correctly predicted the corresponding data for each class, while True Negatives indicate instances where the model accurately predicted data that did not belong to the respective class. Type I Error (False Positive) represents cases where the model incorrectly predicted an item as belonging to a class when it actually does not, and Type II Error (False Negative) refers to instances where the model failed to predict an item as belonging to a class when it actually does.
Confusion matrix for classification
3. Results and discussion
3.1 Characteristics of ultrasonic welds by process parameters
To classify ultrasonic welds based on process parameters, the welds were organized by input energy into lobe areas as shown in Fig. 6, using standards such as appearance, cross-section, T-Peel load, and fracture mode evaluation14). Welds with input energies below 400 J were classified as insufficiently welded, those in the 400-1,000 J range as sufficiently welded, and those exceeding 1,000 J as excessively welded.
Schematic of the lobe area for each ultrasonic welding process parameter
3.1.1 Appearance and cross-section
Table 4 shows the appearance and cross-section characteristics of ultrasonic welds based on process parameters. All insufficient welds were formed under the condition of a welding time of 0.2 s. The appearance of insufficient welds featured shallow indentations, while excessive welds exhibited cracks at the top. Additionally, under conditions where the input energy exceeds 1,000 J, excessive indentation depth was observed, leading to a reduction in the effective thickness and joining area of the joint, as shown in Fig. 7. It was confirmed that when the indentation depth exceeds 0.4 mm, the effective thickness and joining area of the weld decrease. This occurs in the range of input energy exceeding 1,000 J.
Characteristics of ultrasonic welds for each class (insufficient, sufficient, excessive)
Cross-section of the ultrasonic welds showing peaks, valleys, and effective thicknesses
3.1.2 T-Peel load
In the T-Peel load graph shown in Fig. 8, the maximum T-Peel load was 153.55 N at 0.12 MPa, 25 μm, and 0.8 s. However, due to cracks present in the appearance of the weld, it was classified as an overheated weld. Additionally, in all conditions where the input energy exceeds 1,000 J, the presence of cracks at the top and bottom of the foil makes it difficult to accurately predict the load during T-Peel testing. Therefore, the condition with 0.1 MPa, 22 μm, and 0.6 s, which satisfies the appearance and cross-section evaluation, is considered the most optimal. It was observed that as the welding time increased in all conditions, both input energy and load tended to increase, and a significant increase in load was noted when the welding time reached 0.4 s, indicating the formation of adequately joined welds.
T-Peel load of ultrasonic welds
3.1.3 Fracture mode
In Table 5, insufficient welds, which were formed under conditions with input energy below 400 J, exhibited an average T-Peel load of 5.73 N, and all showed interfacial failure. This suggests that there is insufficient input energy to induce adequate plastic deformation in the weld during the ultrasonic welding process. Sufficient welds, formed under conditions with input energy between 400 and 1,000 J, had an average load of 71.68 N and exhibited partial interfacial failure. For conditions where the input energy exceeded 1,000 J, the average load measured was 128.13 N, and tearing or button failure occurred.
Fracture mode of ultrasonic welds
3.2 Evaluation of classification models based on metrics
Table 6 presents the accuracy, precision, recall, and F1-score values for each class (Class 0, 1, 2) of the classification model. The accuracy across the entire dataset was approximately 96.67%. The precision for the model’s predictions was 100% for insufficient welds (Class 0) and overheated welds (Class 2), while it was 94% for sufficient welds (Class 1). The recall for actual insufficient welds was 92%. The F1-scores were all above 0.95, indicating high accuracy across the board.
Metric results for each weld class
A total of 108 data points were classified into Class 0, 1, and 2 based on ultrasonic welding process parameters. The training data showed high accuracy and no misclassification for all classes. In the validation data, two instances of Class 0 were misclassified as Class 1, resulting in a decreased recall of 0.90 for Class 0. As shown in Fig. 9, in the test data, one instance of Class 0 was misclassified as Class 1. Although the classification accuracy for Class 0 was lower compared to Classes 1 and 2, it still achieved 90% performance. It is expected that the accuracy will improve with a larger dataset.
Comparison of confusion matrix for test data
4. Conclusion
In this study, a classification model was developed and evaluated for ultrasonic welds based on the characteristics of welds with varying process parameters, including insufficient welds, sufficient welds, and excessive welds. For this purpose, convolutional neural networks (CNN) and long short-term memory networks (LSTM), which are types of deep learning models, were applied to classify ultrasonic welds. The conclusions are as follows:
1) The evaluation of ultrasonic welds in terms of appearance, cross-section, T-Peel load, and fracture characteristics revealed that insufficient welds exhibited indistinct indentation shapes and inadequate indentation depth. Sufficient welds showed uniform indentation shapes, achieved 102.86 N at 0.1 MPa, 22 μm, and 0.6 s, and exhibited partial interfacial failure. Excessive welds showed cracks in the appearance, exceeded an indentation depth of 0.4 mm, and exhibited tearing and button failure.
2) By using convolutional neural networks to extract local features and long short-term memory networks to learn temporal dependencies, an optimized model for time-series data was developed, achieving an overall accuracy of approximately 97%.
3) It was observed that the accuracy for Class 0 was lower compared to Classes 1 and 2. This is attributed to differences in resolution during power data collection, and it is anticipated that increasing the resolution slightly will enable the model to learn Class 0 more effectively.
