Abstract
Ultrasonic metal welding (UMW) is widely used in industrial applications such as electric vehicle battery manufacturing and automotive body assembly. The joint quality of UMW is sensitive to a variety of undesired yet inevitable process disturbances such as tool condition and material surface condition. Online monitoring and control capabilities are critically needed to quickly detect process anomalies and adaptively adjust process parameters, thus ensuring satisfactory joint quality and reducing process variability. While existing research has developed online quality monitoring methods for UMW, the cost-effectiveness of hardware (sensors, data acquisition equipment, and edge-computing devices) and software (computational efficiency) in a monitoring system has not been investigated systematically. The cost-effectiveness of monitoring also decides the window during which control actions can be executed, thus ultimately influencing the joint quality. This paper presents a systematic study on three factors related to cost-effectiveness: (1) sensor selection, (2) sampling rate of data acquisition, and (3) signal fraction. A method based on discrete wavelet transformation (DWT) is used for automatic feature extraction due to its effectiveness in extracting both time-domain and frequency-domain information with varying signal lengths. We develop a multi-layer perceptron (MLP) classifier to process DWT-generated features and recognize two types of welding disturbances including tool condition and material surface condition. Case studies demonstrate that combining a suitable signal fraction with a subset of sensors leads to comparable performance to using all sensors at full length. Moreover, the interaction between the sampling rate and the signal fraction is investigated. Results confirm the feasibility of building an accurate monitoring system with limited sampling rate and signal fraction, which increases the window for real-time control.
