As manufacturing systems become more automated and interconnected, diagnosing faults and identifying their root cause has become increasingly complex for human operators. Data-driven methods can help prevent costly downtime by leveraging Causal Discovery (CD) to map out how different machine components affect each other, while automated Root Cause Analysis (RCA) tracks down fault origins. However, progress in developing RCA and CD methods is hindered by the lack of real-world datasets that support their joint benchmarking in realistic manufacturing environments. We introduce the causRCA manufacturing dataset to fill this gap. The dataset comprises 170 data recordings from normal operation of a CNC vertical lathe and 100 simulated fault data recordings generated through a hardware-in-the-loop setup that combines a digital twin of the lathe with a physical controller. The dataset includes an expert-validated causal graph connecting the 92 included variables and alarms, serving as ground truth for evaluating both CD and causal RCA methods. We illustrate the versatility of causRCA through exemplary benchmarks that compare supervised RCA methods, unsupervised RCA methods, and CD algorithms on the dataset. Furthermore, we demonstrate its potential for answering research questions regarding causal RCA methods by analyzing how the quality of learned causal graphs affects RCA performance. All data, code, and documentation are publicly available to accelerate research in CD and automated RCA.
Mehling et al. (Thu,) studied this question.