Improving process monitoring of ultrasonic metal welding using classical machine learning methods and process-informed time series evaluation

Abstract

Ultrasonic metal welding (USMW) is an industrially applied joining technology that is highly complex since the weld quality is influenced by numerous factors. The relationships between these factors and the quality remain partly unavailable resulting in a need for improvement in process monitoring and quality management. This work focuses on exploring the relationships between tensile shear strength (TSS) of Cu-sheet welds and process curves from welding machine and additional vibration sensors at sonotrode and anvil. Discovered relationships would enable an improved process monitoring, when valid for a broad parameter range. To ensure the latter, examinations are carried out on a central composite design of experiments (DoE) data set. For the whole data set as well as for single data points, the process curves are examined in detail comprising visualizations and discussions of revealed trends. These trends are related to process physics to clarify their relevance for the TSS. Based on physical process knowledge, more than 700 features are extracted from the curves. The extraction approach is not limited to the present setup and enables a quantitative evaluation of the relation between TSS and process curves. Most important features are derived from the generator power and the anvil vibration. Finally, linear regression as well as multi-layer perceptron regression are used to predict the TSS based on the most relevant features. Comparing the obtained regression results with the reference model, that is the polynomial regression model from standard DoE evaluation, a prediction improvement of nearly 50% is achieved. These results suggest the employed signals as a suited basis for an improved USMW process monitoring.