Bayesian network-based methodology for selecting a cost-effective sewer asset management model | JoVE Visualize

Area of Science:

Civil Engineering
Asset Management
Data Science

Background:

Effective management of sewer infrastructure relies on accurate prediction of structural conditions.
Identifying critical variables for predictive modeling is essential for efficient asset management strategies.
Current methods may involve numerous variables, complicating analysis and prediction.

Purpose of the Study:

To develop and present a methodology using Bayesian networks (BN) for prioritizing and selecting a minimal set of variables for predicting sewer asset structural condition.
To support proactive management strategies through robust and straightforward variable selection.
To validate the methodology's effectiveness using a real-world case study.

Main Methods:

Development of a methodology integrating Bayesian network (BN) models.
Utilization of statistical measures of agreement, specifically Cohen's Kappa coefficient, for evaluating prediction accuracy.
Application of a statistical test (Wilcoxon test) for robust variable selection.
Determination of optimal calibration and validation data subsets (e.g., 80% calibration, 20% validation).

Main Results:

The methodology successfully identified a minimal set of variables (age and diameter) that achieved prediction capacity comparable to models using multiple variables (Cohen's Kappa coefficient = 0.43).
The optimal data split for calibration (80%) and validation (20%) was identified, enhancing predictive capacity with low variation.
Focusing the model on pipes in critical and excellent conditions significantly increased prediction success (Cohen's Kappa coefficient increased from 0.2 to 0.43).

Conclusions:

The proposed BN-based methodology provides a robust and efficient approach for selecting minimal variables to predict sewer pipe structural conditions.
This approach simplifies predictive modeling, enabling more effective and proactive sewer asset management.
The findings demonstrate the potential for significant improvements in prediction accuracy by focusing on specific condition states and key variables.