Anomaly Detection in Optical Monitoring of Laser Beam Welding

Abstract

Robotized laser beam welding is one important process in manufacturing, offering efficient welding while minimizing the heat input. Nonetheless, this method is sensitive to various deviations, including fixture problems, heat-induced distortions, and inaccuracies in tool handling. Such deviations can lead to significant defects like lack of fusion, particularly when welding square butt joints without gaps. Detecting these defects through visual inspection or non-destructive methods is challenging. To address this, real-time monitoring and automatic intervention are necessary. One effective sensor for monitoring laser beam welding is the photodiode, which captures optical emissions from the process. Research has demonstrated correlations between these emissions and process stability. Photodiodes are cost-effective and easily integrated into welding tools, making them ideal for industrial applications. However, the challenge lies in analyzing the output signals and defining thresholds for identifying deviations from normal conditions. Thus, there’s a need for an automated method to set threshold values based on measured data. Machine learning offers a solution, particularly through supervised, unsupervised, or semi-supervised methods. Supervised machine learning requires labeled data, involving time-consuming experiments with nominal and deviating cases, making it less feasible for industrial setups. This paper suggests using unsupervised learning for anomaly detection, relying solely on data from nominal welding cases for model training. In this approach, a model is trained using photodiode data from a single nominal weld case and subsequently tested on data collected during experiments involving laser beam offsets during welding. The results demonstrate the promise of this method for monitoring closed square butt-joint laser beam welding, even with limited training data from nominal cases.