A Multi-task Interpretable Few-shot Learning Framework for Ultrasonic Welding Quality Recognition

Abstract Ultrasonic metal welding (UMW) is a high-efficiency, low-thermal-impact solid-state bonding technique that has been widely employed in battery manufacturing, the automotive industry, and related fields. However, its performance is highly sensitive to process anomalies and variations introduced by new process configurations, which may result in significant degradation of weld quality and substantial production losses. Consequently, developing online monitoring approaches with strong generalization capability has become essential.Conventional methods often depend on large volumes of labeled data and lack adaptability to unseen process conditions. To address these limitations, this study proposes a novel multi-task interpretable few-shot learning framework, termed MXFSL. The framework integrates DE-VMD-based feature extraction, multi-task collaborative meta-learning, and a multi-scale SHAP interpretation module to enable efficient anomaly detection and feature attribution under limited data scenarios.Experimental results demonstrate that MXFSL achieves superior classification performance and robustness in few-shot settings. Furthermore, it identifies critical features contributing to welding quality across different stages, offering valuable insights for intelligent monitoring and adaptive process optimization in UMW applications.