An end-to-end online quality prediction system for ultrasonic metal welding based on deep learning

Ultrasonic metal welding (UMW) is an important joining technology that is widely used in industry. In many UMW applications, there is a strong need for predicting joint quality quickly, reliably, and non-destructively. State-of-the-art quality assessment methods such as destructive tensile testing and quality monitoring cannot meet the high requirements in industrial-scale production. This thesis proposes a novel end-to-end online prediction algorithm for UMW based on deep learning that offers various benefits, including superior quality prediction, less reliance on prior knowledge of UMW processes (e.g., tool conditions), and not involving tedious data preprocessing and feature engineering. The effectiveness of the proposed method is shown using real-world data generated from a UMW process. A comparative case study is presented to compare three data fusion strategies (early fusion, middle fusion, and late fusion) and traditional feature engineering-based methods. The results show that the proposed end-to-end quality prediction system outperforms traditional methods. In addition, the middle fusion strategy achieves the best prediction performance