causRCA: Real-World Dataset for Causal Discovery and Root Cause Analysis in Machinery

causRCA: Real-World Dataset for Causal Discovery and Root Cause Analysis in Machinery causRCA is a collection of time series datasets recorded from the CNC control of an industrial vertical lathe. The datasets comprise real-world recordings from normal factory operation and labeled fault data from a hardware-in-the-loop simulation. The fault datasets come with labels for the underlying (simulated) cause of the failure, a labeled diagnosis, and a causal model of all variables in the datasets. The extensive metadata and provided ground truth causal structure enable benchmarking of methods in causal discovery, root cause analysis, anomaly detection, and fault diagnosis in general. Use Cases & Applications Causal Discovery: Benchmark learned causal graphs against an expert-derived causal graph. Supervised Root Cause Analysis: Train and test models on labeled diagnosis for different fault scenarios. Unsupervised Root Cause Analysis: Identify manipulated variables in different fault scenarios with known ground truth. Data & File Overview data/ ┣ real_op/ ┣ dig_twin/ ┃ ┣ exp_coolant/ ┃ ┣ exp_hydraulics/ ┃ ┗ exp_probe/ ┣ expert_graph/ ┗ README_DATASET.md The data folder contains: real_op/: CSV files with time series data from normal operation. dig_twin/: Data from the digital twin experiments. Each group (coolant,hydraulics,probe) contains a causal subgraph as ground truth, different fault scenarios and multiple runs per scenario: exp_coolant/: Coolant system faults exp_hydraulics/: Hydraulic system faults exp_probe/: Probe system faults expert_graph/: GML and interactive HTML file with the expert-derived causal graph and lists of nodes and edges. README_DATASET.md: Dataset description Datasets summary (Sub-)graph #Nodes #Edges #Datasets normal #Datasets Fault #Fault Scenarios #Different Diagnoses #Causing Variables Lathe (Full graph) 92 104 170 100 19 10 14 --Probe 11 15 170 34 6 3 2 --Hydraulics 17 18 170 41 9 5 6 --Coolant 15 10 170 25 4 2 6 --(Other Vars) 49 61 170 - - - - *datasets from normal operation contain all machine variables and therefore all subgraphs and their respective variables within it. Methodological Information Real Operation Data (real_op) Data were recorded through an OPC UA interface during normal production cycles on a vertical lathe. These files capture baseline machine behavior under standard operating conditions, without induced or known faults. Digital Twin Data (dig_twin) A hardware-in-the-loop digital twin was developed by connecting the original machine controller to a real-time simulation. Faults (e.g., valve leaks, filter clogs) were injected by manipulating specific twin variables, providing known ground-truth causes. Data were recorded via the same OPC UA interface to ensure consistent structure. Known limitations Data was sampled via an OPC UA interface. The timestamps only reflect the published time of value change by the CNC and do not necessarily reflect the exact time of value changes. Consequently, the chronological order of changes across different variables is not strictly guaranteed. This may impact time-series analyses that are highly sensitive to precise temporal ordering. Methods for Processing see the causRCA GitHub Repository Acknowledgements The authors gratefully acknowledge the contributions of: KAMAX Holding GmbH & Co. KG for providing real production data from the vertical lathe. Schuster Maschinenbau GmbH for supporting the digital twin development with knowledge and the PLC project. ISG Industrielle Steuerungstechnik GmbH for developing the digital twin implementation. SEITEC GmbH for hosting the hardware-in-the-loop setup and developing the OPC UA data recording solution. Declaration of GenAI and AI-assisted Technologies During the preparation of the dataset, the author(s) used generative AI tools to enhance the dataset's applicability by structuring data in an accessible format with extensive metadata, assist in coding transformations, and draft description content. All AI-generated output was reviewed and edited under human oversight, and no original dataset content was created by AI.