Multi-step time-to-failure predictions in water pipelines using feature engineering and cascading ensembles

Area of Science:

• Civil Engineering
• Water Resource Management
• Data Science

Background:

• Accurate prediction of water pipeline failure timings is vital for effective maintenance and rehabilitation strategies in water distribution networks.
• Existing models struggle to integrate failure history and predict sequential failures, limiting their practical application.
• The accuracy of current time-to-failure models is often compromised by insufficient investigation into significant failure predictors within historical data.

Purpose of the Study:

• To develop advanced models for predicting water pipeline failure timings with enhanced accuracy.
• To address the limitations of existing models in incorporating failure history and predicting sequential failures.
• To identify and leverage the most significant predictors of pipeline failure from historical and dynamic data.

Main Methods:

• Development of a customized weather index and other interaction features to capture time-based deterioration impacts.
• Automated feature selection using Genetic Algorithms (GA) to optimize model performance.
• Implementation of a cascading ensemble model for predicting multiple sequential failures (first, second, third) in individual pipelines.
• Validation using both hold-out and out-of-sample testing.

Main Results:

• Feature engineering and GA-based feature selection led to a 20-50% increase in model performance.
• The developed models achieved Mean Absolute Error (MAE) ranges of 1.4-0.5 for feature selection optimization.
• The cascading ensemble model demonstrated superior performance (MAE: 0.8-1.1) compared to alternative multi-output models for sequential failure prediction.
• A web-based application was developed to demonstrate the practical utility of the novel modeling approach.

Conclusions:

• The study presents a novel modeling regime for high-performance failure timing prediction in water pipelines.
• The approach offers micro-level analysis of pipe sections, providing insights into complex feature interactions and deterioration rates.
• The developed models and methods significantly improve the prediction of individual pipeline failures, supporting better infrastructure management.