Cluster-Based Approach for Visual Anomaly Detection in Multivariate Welding Process Data Supported by User Guidance

Welding robots are essential in modern manufacturing as they automate hazardous welding tasks, improving productivity and safety while reducing costs. However, a significant portion of the total part costs comes from the manual visual inspection and the rework of robot-welded seams, underlining the importance of process optimization. Production sites are increasingly digitalized, using systems to track and manage production processes, plan resources, and collect production process data. Utilizing this data, welding engineers face the challenge of analyzing extensive time series data to gain actionable insights. The complexity and volume of the data make it challenging to identify problems, while missing ground truth and labels make unsupervised approaches, such as anomaly detection for short-term issues and clustering for long-term trends, necessary. To ensure that our research fits the specific needs of welding engineers, we conducted a design study with subject matter experts from the industry. Based on the design study, we introduce a visual analytics approach to support domain experts in analyzing welding data, addressing the challenge of examining multiple time series datasets recorded from different welding robots that produce multiple seams on different components within a production line. The interactive tool integrates advanced visualization techniques in a human-in-the-loop approach to allow domain experts to identify, explore, and interpret anomalies and clusters. It implements directing guidance to support users with navigating and focusing on meaningful patterns in data. A pair analytics user study assessed the prototypes' capabilities in hypothesis generation and examined how well users could learn and utilize the system efficiently. The study presents examples of findings, demonstrating how domain expert participants utilize the visual analytics tool to reveal patterns, leading to potentially improved decision-making and operational efficiency. We conclude the article with possible future work directions for researchers aiming to refine our tool's capabilities.