Customised-sampling approach for pipe failure prediction in water distribution networks

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Imbalanced datasets pose challenges for machine learning (ML) models in predicting failures in Water Distribution Networks (WDNs).
• Traditional methods struggle to accurately identify rare failure events due to skewed data distributions.
• Effective failure prediction is crucial for maintaining WDN integrity and operational efficiency.

Purpose of the Study:

• To develop and evaluate a novel methodology for addressing imbalanced class data in WDN pipe failure prediction.
• To investigate the impact of various data balancing strategies on the performance of ML models.
• To identify optimal configurations of under-sampling, over-sampling, and class weighting for improved predictive accuracy.

Main Methods:

• Utilized under-sampling, over-sampling, and class weighting techniques to rebalance imbalanced WDN datasets.
• Constructed pipe failure prediction models using these adjusted datasets at various levels, including non-balance points.
• Evaluated model performance using F1-score and Area Under the Receiver Operating Characteristic Curve (AUC-ROC).

Main Results:

• Under-sampling above the balance point resulted in the highest F1-score.
• Over-sampling below the balance point demonstrated optimal performance.
• Applying class weights with lower values than the balance point proved effective.
• Combining over-sampling and under-sampling at different ratios, followed by class weighting, yielded the most effective predictive model.

Conclusions:

• Data balancing techniques are critical for improving failure prediction in imbalanced WDN datasets.
• A hybrid approach combining targeted over-sampling, under-sampling, and class weighting offers superior predictive performance.
• The findings provide valuable insights for developing robust WDN failure prediction systems.