Hybrid Digital Twin Dataset – Detergent Plant
Description
This dataset contains anonymized batch-level data collected from SCADA, MES, and ERP systems in a detergent manufacturing plant. It includes variables related to raw material volumes, batch durations, reactor allocation, and laboratory analysis time. The dataset was used to train and evaluate predictive models integrated into a hybrid Digital Twin architecture for process optimization under Industry 5.0 principles.
Files
Steps to reproduce
Dataset Overview: The dataset consists of batch-level records from a detergent manufacturing plant, extracted from SCADA, MES, and ERP systems. Each row represents a production batch, and columns include variables such as raw material volumes, loading time, batch version, reactor ID, and analysis time. Variable Dictionary: Refer to the file variables_dictionary.pdf for a complete description of each column and its data type. Data Cleaning: All sensitive identifiers were removed or anonymized. Missing values were handled using domain knowledge or removed if non-essential. Modeling Workflow (Described in Manuscript): Predictive tasks were structured around supervised learning, including batch loading time (regression), reactor allocation (classification), and quality proxy (binary classification). Preprocessing included normalization, categorical encoding, and time-series alignment. Models used included Random Forest, Gradient Boosting, and Ridge Classifier. Hybrid Integration: While code is not shared due to intellectual property protection, the manuscript details the structure of the hybrid models, which combine ML predictions with theoretical values derived from Ordinary Differential Equations (ODEs). Reproduction Tips: Users can reconstruct similar models using Python (e.g., Scikit-learn, XGBoost) with the provided data. Suggested targets: loading_time_minutes, reactor_id, analysis_time_exceeded (binary proxy). Suggested features: batch_version, ingredient_volumes, reactor, total_mass, formulation_ID. Limitations: Reproduction of exact results may not be possible without the original plant codebase and sensor configurations. However, similar workflows can be reproduced for benchmarking or transfer learning.
Institutions
- Universidade Federal do Vale do Sao Francisco