Distillation Column Soft-Sensor

Machine Learning-based "Virtual Soft-Sensor" trained on high-fidelity simulation data representing 450 hours of industrial operation. It predicts ethanol purity in real-time, robustly handling noise and outliers to replace expensive hardware analyzers.

---

📋 Project Overview

In industrial chemical processing, measuring product purity (e.g., Ethanol concentration) in real-time is often challenging due to the high cost and latency of on-line hardware analyzers.

This project implements a Virtual Soft-Sensor for a binary distillation column. By leveraging a Random Forest Regressor, the system infers the distillate product quality instantly using readily available process variables (Temperatures, Pressures, and Flow Rates).

Key Capabilities:

• 🚀 Real-Time Prediction: Instantly estimates Ethanol Purity based on sensor inputs.
• 📊 Process Insights: Flags production as ON-SPEC (≥98%) or OFF-SPEC and provides reasoning.
• 🌡️ Visual Analytics: Visualizes the column temperature profile to detect anomalies like flooding or drying.
• 💡 Actionable Recommendations: Suggests control actions (e.g., "Increase Reflux Ratio") to operators.

---

🧬 About the Dataset

The model is trained on a synthetic dataset generated from a rigorous mathematical model of a binary distillation column. To ensure industrial relevance, the data includes atypical features such as sensor noise, outliers, and missing values, simulating real-world operating conditions.

• Source: Mathematical Simulation of a Distillation Column
• Duration: 450 Hours of continuous operation
• Sampling Rate: Every 0.1 hours (6 minutes)
• Total Samples: ~4,500 data points

Feature Dictionary

The dataset consists of 22 features capturing the thermodynamic state of the column:

Variable Type	Feature Name	Description	Unit
Control	Pressure	Column operating pressure	bar
State	Temp_Tray_2 ... Temp_Tray_15	Temperatures at various tray levels	Kelvin (K)
Flow	Liquid_Flow	Internal liquid flow rate	kmol/h
Flow	Vapor_Flow	Internal vapor flow rate	kmol/h
Flow	Distillate_Flow	Top product flow rate	kmol/h
Flow	Bottoms_Flow	Bottom product flow rate	kmol/h
Flow	Feed_Flow	Input feed flow rate	kmol/h
Target 🎯	Ethanol_Mol	Molar concentration of ethanol	Mol Fraction

📚 Citation

If you find this approach or data useful, please acknowledge the original dataset creators:

Cote-Ballesteros, J. E., Grisales Palacios, V. H., & Rodriguez-Castellanos, J. E. (2022). Un algoritmo de selección de variables de enfoque híbrido basado en información mutua para aplicaciones de sensores blandos industriales basados en datos. Ciencia E Ingeniería Neogranadina, 32(1), 59–70. https://doi.org/10.18359/rcin.5644

🛠️ Tech Stack

• Core: Python 3.10+
• Machine Learning: Scikit-Learn, Pandas, NumPy
• Web Framework: Flask (Backend), Jinja2
• Frontend: HTML5, TailwindCSS, Chart.js
• MLOps: Joblib for model serialization

---

📊 Model Performance & Insights

Model:
Random Forest Regressor

• n_estimators = 150
• max_depth = 15

📈 Performance Metrics

• R² Score: > 0.95 (test set)
• MAE: ± 0.005

---

🏆 Optimum Operating Conditions (Golden Batch)

Based on data-driven optimization, the system reaches ~89% product purity under the following operating conditions:

Parameter	Value	Unit
Reflux Flow (L)	1040.0	kmol/h
Feed Flow (F)	650.0	kmol/h
Vapor Flow (V)	1300.0	kmol/h
Distillate (D)	260.0	kmol/h
Bottoms (B)	390.0	kmol/h
Bottom Temperature (T14)	360.0	K
Top Temperature (T2)	350.8	K

---

📌 Note:
These conditions represent the Golden Batch, indicating maximum efficiency and optimal purity according to the trained model.

📂 Project Structure

distillation-column-monitor-mlops/
├── data/
│   ├── raw/                 # Original CSV data (distillation_data.csv)
│   └── processed/           # Cleaned data for training
├── models/
│   └── final/               # Trained model artifacts (.pkl)
├── src/
│   ├── templates/           # HTML Frontend (index.html)
│   ├── app.py               # Flask Application Entry Point
│   ├── config.py            # Configuration settings
│   ├── data_prep.py         # Data Cleaning & Feature Engineering
│   ├── inference.py         # Prediction & Business Logic
│   └── train.py             # Model Training Script
├── requirements.txt         # Project Dependencies
└── README.md

🚀 Installation & Setup

1. Clone the Repository Bash

git clone https://github.com/Kochipek/Distillation-Column-Soft-Sensor-mlops.git
cd Distillation-Column-Soft-Sensor-mlops

2. Create a Virtual Environment

It is recommended to use a virtual environment to manage dependencies.

macOS / Linux:

python3 -m venv venv
source venv/bin/activate

Windows:

python -m venv venv
venv\Scripts\activate

3. Install Dependencies

pip install -r requirements.txt

4. prepare and train data

python src/data_prep.py
python src/train.py

5. Run application

python src/app.py