Abstract Ultrasonic metal welding is a widely used process for joining metals to create electrical connections using high-frequency mechanical vibrations. Despite its industrial relevance, achieving consistent joint quality remains challenging and requires a deeper understanding of the relationship between process parameters and joint quality. The challenge here is identifying the process signals that significantly correlate with the joint quality. Due to the nature of the process, manual evaluation of such signals is ineffective, leading to an increased focus on implementing statistical or machine learning-based models for predicting joint quality. This study proposes a method to improve the accuracy and interpretability of predicting joint quality using machine learning models. Welding experiments involving various machine internal and external sensors are conducted. The sensors’ measurement data are recorded synchronously as time series. Since most machine learning models require scalar input, informative features must be extracted from the time series data. Various statistical and signal-based features can be extracted from the entire time series. To distinguish between the influences of the different process phases and their overall behavior, the signals are divided into segments. Although equal-length segmentation is simple, it overlooks signal characteristics such as steep transitions or prolonged steady states. Change point detection enables segmentation based on signal behavior, allowing for more relevant feature extraction. Moreover, the detected segments can be associated with known process phases, linking data-driven insights with domain knowledge. As a result, both the predictive performance and interpretability of the ML models improve, allowing for clearer associations between process phases and quality outcomes.
Stockemer et al. (Thu,) studied this question.