Machine learning as a comparative tool to determine the relevance of signal features in laser welding

Abstract In laser welding, real-time process information is required for quality assurance and process control. The challenge on sensor technology and processing is determined by dynamic process behavior in the interaction zone. The dimensionality of spatial process information however can be used to determine different process conditions and quality faults. In this paper, a laser welding process is coaxially observed using a high-speed mid-infrared-camera in combination with near-infrared CMOS-camera. The signals are evaluated by extracting geometrical and statistical features. These features are subsequently used by machine learning techniques to predict quality-relevant process variables. The work identifies and compares the importance of the extracted features based on their prediction performance. This importance is used to reassess the sensors and underlying signal analyses in order to enhance the relevance of the acquired process information.