Batch Distillation Data for Developing Machine Learning Anomaly Detection Methods

Area of Science:

• Chemical Engineering
• Data Science
• Process Control

Background:

• Machine learning (ML) shows promise for anomaly detection (AD) in chemical processes.
• Development of ML-based AD methods is limited by the scarcity of public experimental data.

Purpose of the Study:

• To address the data gap for ML-based AD in chemical processes.
• To generate a comprehensive experimental dataset for training and validating ML models.
• To enable the development of interpretable and explainable AD methods.

Main Methods:

• Established a laboratory-scale batch distillation plant for data generation.
• Conducted 119 experiments with varying operating conditions and mixtures, including induced anomalies.
• Collected time-series sensor/actuator data, measurement uncertainty, NMR spectroscopy, video, and audio recordings.

Main Results:

• Generated an extensive, structured dataset with fault-free and anomalous experiments.
• Included detailed metadata, expert annotations, and an anomaly ontology.
• Dataset is publicly available, facilitating ML-based AD research.

Conclusions:

• The new dataset supports the advancement of ML-based AD in chemical processes.
• Enables development of interpretable, explainable, and mitigable AD solutions.
• Promotes further research in data-driven process monitoring and control.