CIPCaD-Bench: Continuous Industrial Process datasets for benchmarking Causal Discovery methods

Industrial process optimization and fault diagnosis have long depended on expert-driven models and heuristic analysis. Traditional methods, such as root cause analysis and control charts, often lack scalability and adaptability to complex, high-dimensional data. Recent advances in causal discovery algorithms—like LiNGAM, GES, PC, and NOTEARS—offer promising avenues for automating causal inference from observational data. However, their validation has been limited by the scarcity of publicly available, realistic datasets. Synthetic datasets, while controlled, do not fully capture the dynamics of real industrial environments. Conversely, real datasets are often proprietary, making benchmarking difficult. This gap impedes progress in deploying causal inference tools in industry, where accurate, scalable, and interpretable models are crucial for safety, efficiency, and regulatory compliance.

The core challenge lies in establishing reliable causal relationships within complex, noisy, and high-dimensional industrial data. Existing datasets are either synthetic or proprietary, lacking standard benchmarks for algorithm comparison. Moreover, real-world data contain confounders, external influences, and measurement uncertainties, complicating causal inference. Without standardized benchmarks, it is difficult to objectively evaluate and improve algorithms for industrial applications. The problem is further exacerbated by the need for algorithms to balance accuracy, computational efficiency, and robustness in operational settings. Addressing these issues requires publicly accessible, representative datasets and a comprehensive evaluation framework.

This work introduces two innovative datasets: one simulated via Tennessee Eastman, capturing process mechanics with known ground truth, and one from a real food manufacturing plant, reflecting authentic process variability. The datasets include multiple ground truths to evaluate different aspects of causal discovery. A multi-metric benchmarking protocol assesses algorithms on accuracy (TP, FDR), structural similarity (SHD), and computational efficiency, providing a holistic performance view. The systematic comparison of algorithms—LiNGAM, GIES, PC, FCI, NOTEARS—across these datasets reveals their strengths and limitations, guiding practical algorithm selection. The ground truth design for real data incorporates plausible causal models, bridging the gap between synthetic and real-world validation.

• �� Data collection: Simulated data from Tennessee Eastman process, real data from ultra-processed food plant, both preprocessed via normalization and sliding window filtering.
• �� Variable selection: 33 variables in TE, 17 in food plant, covering key process and manipulated parameters.
• �� Ground truth construction: Based on process mechanics for TE, and process flow for real data, with two models for the latter to account for internal dependencies.
• �� Algorithm implementation: Running LiNGAM, GIES, PC, FCI, NOTEARS, with hyperparameter tuning for fairness.
• �� Performance metrics: TP, FDR, SHD, PR, RE, evaluated across datasets.
• �� Comparative analysis: Statistical tests and visualizations to interpret differences in structure recovery, robustness, and speed.

Experiments involved applying multiple algorithms to both datasets, with consistent preprocessing and parameter tuning. The TE dataset, with known ground truth, allowed direct accuracy assessment. The real food data tested algorithm robustness in complex, noisy environments. Multiple runs ensured result stability. Hyperparameters such as regularization strength and learning rate were optimized via grid search. The evaluation focused on structure accuracy, false positive/negative rates, and computational time. Sensitivity analyses examined the impact of data preprocessing choices. The experiments aimed to identify algorithms best suited for industrial applications, balancing accuracy and efficiency.

LiNGAM and GIES achieved over 85% true positive detection with minimal SHD in the TE dataset, indicating high structural fidelity. In real data, PC and FCI maintained FDR below 0.2, demonstrating robustness amid process variability. Speed-performance trade-offs were evident: LiNGAM was accurate but computationally intensive, while PC was faster with slightly lower accuracy. The multi-metric evaluation highlighted that no single method dominates across all criteria, emphasizing the need for context-specific algorithm choice. The results validate the datasets' utility for benchmarking and guide future algorithm development.

The datasets and evaluation framework support fault diagnosis, process optimization, and causal model deployment in industry. Practitioners can benchmark new algorithms, tune parameters, and develop real-time causal inference tools. The work facilitates integration of causal discovery into industrial control systems, enabling predictive maintenance and autonomous decision-making. Long-term, these tools could lead to fully automated, self-optimizing manufacturing processes, reducing downtime and improving safety.

While comprehensive, the datasets are limited by sampling frequency and scope, potentially missing fast or subtle causal effects. High computational costs of some algorithms hinder real-time deployment. The causal directions inferred may be uncertain without experimental intervention, especially in complex systems. Future work should incorporate interventional data, improve scalability, and explore adaptive algorithms to address these challenges.