1D CNN-LSTM Based on Quality Classification in Ultrasonic Metal Welding of Multi-Layered Foil Stacks

Ultrasonic metal welding (UMW) is widely utilized for joining multi-layered current collector foil stacks in lithium-ion battery manufacturing, where online joint quality monitoring is critical to prevent defects. In this study, the linear ultrasonic welding of a 60-layer pure Cu foil stack (8 µm thickness each) was investigated by establishing an in situ multi-modal sensing system and a deep learning-based quality classification framework. A linear variable differential transformer (LVDT) sensor and an ultrasonic power meter were synchronized to simultaneously capture mechanical horn displacement and electrical power signals across sequential welding stages. Experimental characterization revealed distinct phenomenological regimes in the acquired profiles, with the vertical displacement showing a direct correlation with the knurl penetration and weld indentation depth. Based on the delivered welding energy, the joint quality was classified into three distinct regimes: under-welding (<500 J), normal welding (500–900 J), and over-welding (>900 J). A two-branch hybrid 1D Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture was developed to effectively learn both local transient patterns and long-term temporal dependencies from the preprocessed time-series signals. The proposed classification model achieved an overall prediction accuracy of 98% (macro F1-score of 0.98) on the held-out test set, demonstrating high classification performance for individual welding signals within the investigated process domain. These results demonstrate the feasibility of using multi-sensor in situ data combined with hybrid deep learning models for intelligent, real-time quality assurance in multi-layered Cu foil (8 µm) stack joining.