Ultrasonic metal welding is a solid-state joining technique with a broad range of industrial applications, yet process fluctuations remain common. These inconsistencies are frequently linked to an incomplete understanding of the complex interactions between tooling and joining partners. To address this limitation, the present study derives characteristic features from airborne acoustic emissions that correlate with the thermomechanical phases of the welding process, thereby aiming to deepen fundamental process understanding and support future monitoring strategies. Copper sheets (CW004A) were welded at 20 kHz while vibrations of the horn and upper joining partner were captured simultaneously using laser vibrometry. The process phases were identified by detecting inflection points in the relative velocity profiles, which served as the ground truth for process phase segmentation. To validate these phase boundaries, shear tensile tests and microscopic analyses of fracture surfaces were performed at selected welding stages. Airborne acoustic emissions were recorded using a laser-optical microphone (10 Hz–1 MHz) and processed to extract a structured feature set comprising time-domain descriptors (zero-crossing rate, root mean square, entropy, skewness, kurtosis), frequency-based features (spectral peaks, roll-off, flatness, centroid), and generator-signal-based parameters. Building on these features, machine learning and deep learning methods were applied to detect distinct welding phases. A monotonic Hidden Semi-Markov Model (HSMM) with Viterbi decoding was further employed to ensure temporally consistent segmentation. The resulting phase detection accuracy closely matches the vibrometric ground truth. Overall, the findings advance the understanding of acoustic emissions in USMW and demonstrate the potential of airborne sensing for non-invasive monitoring.
Ehlich et al. (Wed,) studied this question.