Wastewater BOD Prediction System (MLOps & Compliance)

Overview

This project implements an end-to-end MLOps Platform for real-time wastewater quality monitoring. It deploys a Soft Sensor (Machine Learning Model) to estimate Biological Oxygen Demand (BOD), a critical metric that typically requires a 5-day laboratory test ($BOD_5$).

By correlating real-time physical sensors (pH, Conductivity, COD, Flow) with historical biological data, this system provides instant feedback to plant operators, enabling proactive process control.

Key Objectives

1. Production Engineering: Modular, refactored code structure suitable for Docker/Kubernetes.
2. Observability: Real-time tracking of Data Drift (Salinity spikes) and System Latency.
3. Regulatory Compliance: Full transparency documentation (MODEL_CARD.md) aligned with the EU AI Act.

---

Project Structure

wastewater_project/
├── api/                   # FastAPI Deployment & Monitoring Middleware
│   └── main.py            # Serves the model + /monitoring/stats endpoint
├── config.yaml            # Central configuration (Model types, Paths)
├── data/                  # Data storage (Raw & Processed)
├── docs/                  # Sphinx Documentation
├── models/                # Serialized Models (.joblib)
├── notebooks/             # R&D and Exploratory Data Analysis
├── src/                   # Production Source Code (Refactored)
│   ├── data_loader.py     # Robust data ingestion
│   ├── features.py        # Feature Engineering (Cyclical Time)
│   ├── model.py           # Model Factory (RandomForest/SVM)
│   └── main.py            # Automated Training Pipeline
├── MODEL_CARD.md          # Regulatory Transparency Document
├── report.md    # Executive summary
└── README.md              # You are here

Getting Started

1. Installation

Clone the repository and install dependencies:

pip install -r requirements.txt

(Dependencies include: fastapi, uvicorn, pandas, scikit-learn, pyyaml, mlflow)

2. Train the Model (Automated Pipeline)

Run the MLOps pipeline to process data, engineer features, and train the model. This reads settings from config.yaml.

python src/main.py

• Output: Saves a trained model to models/bod_predictor.joblib.

• Metrics: RMSE ~45 mg/L (Logged to MLflow).

3. Launch the API (Deployment)

Start the FastAPI server to serve predictions and monitor health.

uvicorn api.main:app --reload

• API Docs: http://127.0.0.1:8000/docs

• Monitoring Dashboard: http://127.0.0.1:8000/monitoring/stats

---

Monitoring & Observability

The system includes custom Middleware to track production health. Access the /monitoring/stats endpoint to view:

• Latency: Average inference time (ms).

• Drift Detection: Monitors input Conductivity and COD.

• Alerts: Flags "Shock Load" events if predictions exceed 400 mg/L.

---

Regulatory Compliance (EU AI Act)

This project adheres to Article 13 (Transparency) of the EU AI Act for High-Risk AI Systems.

---

Feature Engineering Highlight

To handle the biological seasonality of wastewater without creating discontinuities at year-end, we implemented Cyclical Time Encoding:

# Converting linear "Month" into continuous coordinates
df['sin_month'] = np.sin(2 * np.pi * df['Month'] / 12)
df['cos_month'] = np.cos(2 * np.pi * df['Month'] / 12)

This ensures the model understands that December (12) and January (1) are neighbors.

---

License

MIT License - Open for educational and operational use.