Principal Component Imagery for the Quality Monitoring of Dynamic Laser Welding Processes

A popular technique to monitor laser welding processes is to record laser-induced plasma radiation with a highspeed camera. The recorded sequences are analyzed using pattern recognition systems. Since the raw data are too high dimensional to allow for an efficient learning, dimension reduction is necessary. The most common technique for dimension reduction in laser welding applications is to use geometric information of segmented objects. In contrast, we propose to adapt ideas from face recognition and to employ appearance-based features to describe the relevant characteristics of the recorded images. The classification performance of geometric and appearance-based features is compared on a representative data set from an industrial laser welding application. Hidden Markov models are used to capture the temporal dependences and to perform the classification of unlabeled sequences into an error-free and an erroneous class. We demonstrate that a classification system based on appearance-based features can outperform geometric features.