Physics-informed ensemble learning for online joint strength prediction in ultrasonic metal welding

Ultrasonic metal welding (UMW) is a promising solid-state joining technology that enables innovative and sustainable manufacturing. Despite possessing numerous advantages, UMW has a narrow operating window and is susceptible to both internal and external disturbances. As such, industrial scale UMW production calls for efficient, effective, and non-destructive joint quality assessment. To this end, this paper develops a novel hierarchical physics-informed ensemble learning (PIEL) framework that uses both physical knowledge and online sensing data for accurate online prediction of UMW joint strength. The PIEL framework decomposes the joint strength variability into a physics-informed global trend and a data-driven residual. The global trend is attributed to controllable or measurable welding conditions and is typically perceived as a large-scale variability. The data-driven residual is observed as a small-scale component when identical welding conditions are applied and can be captured by online sensing data. Drawing on this decomposition, hierarchical prediction models can be established to simultaneously account for both types of variabilities. As an essential component of the PIEL methodology, a highly efficient feature extraction procedure is developed using discrete wavelet transformation (DWT). The DWT-based feature extraction procedure is able to automatically extract key low-dimensional information from high-dimensional sensing signals, thus offering improved efficiency and effectiveness compared to conventional feature engineering approaches. Two real-world case studies with distinct physical setups are presented to demonstrate the effectiveness of the PIEL framework. The first case study investigates the influence of tool degradation and uses a dataset consisting of 200 welding cycles generated under four tool conditions. In the second case study, welding parameters including time, amplitude, and pressure are varied to generate 240 welds. When compared against multiple state-of-the-art baseline methods, PIEL consistently achieves superior prediction accuracy. Further, it is shown that by integrating physical knowledge, PIEL can effectively avoid overfitting and achieve excellent modeling robustness and data efficiency. While developed in the context of UMW, the PIEL framework is readily extensible to various other manufacturing processes.