Hackathon solution by team Mr. Sister. An interactive maintenance-support system for oil-and-gas liquid pipeline networks that combines a physical flow model, a pipe-section model, and machine learning.
For every section of the network the system reports:
- Accumulated corrosion-driven wear.
- Remaining service life.
- Current risk of use, from a predictive model trained on historical downtime.
- Forecast risk growth for that section.
- Recommendations for changing operating regimes to extend service life, including what-if scenarios.
The network graph is rendered interactively with intuitive failure-probability markers. New data can be added through the interface, and the system is designed to be adapted to a live monitoring feed.
Extensive data analysis, many tested hypotheses, and several machine-learning models at different levels of interpretability. Model quality on 3-fold cross-validation, measured by F1, precision and recall, ranges from 0.91 to 0.99.
Pipeline operating-regime telemetry: pressure, temperature, fluid flow, vibration.
- Prepare the data for model training.
- Analyse the data and identify the features that drive pipeline failures.
- Build and train a machine-learning model for failure prediction.
- Evaluate model quality and run acceptance testing.
- Build a user interface for visualising and monitoring pipeline condition.
Python · Jupyter Notebook · pandas, numpy, scikit-learn · matplotlib, seaborn · CatBoost · Flask
- Install Python and Git.
- Clone the repository.
pip install -r requirements.txt- Run the notebooks for data processing and model training.
- Start the web interface:
python main.py
| Name | Role | Contact |
|---|---|---|
| Vadim Olennikov | Data analysis and product | t.me/LTDigor |
| Sergey Vandanov | Machine learning and software engineering | t.me/rapid76 |
| Konstantin Baltsat | Machine learning | t.me/baltsat |