Skip to Main Content
The importance of laser welding in industry increases. Many welds have high-quality demands, and one possibility to satisfy the quality requirements is to monitor the welding process with high-speed cameras. Laser welding is a highly dynamic process; it is therefore challenging to distinguish between normal process fluctuations and abnormal error events in the recorded sequences. This paper investigates a novel classification method to automatically analyze the recorded welding sequences and robustly find the abnormal error events. To our knowledge, it is the first time that a framework to detect and track sputters in welding sequences is proposed and evaluated. To achieve a high usability of the classification algorithm, in the training phase, the user only needs to mark suspicious sequences but does not need to label individual frames within the sequences. The framework is tested on two challenging data sets from real welding processes. The results show that the material particles can be tracked accurately. On a sample data set, the new approach finds all erroneous welds with a small false-positive rate and outperforms previously developed methods.