Abstract:

To address the insufficient robustness of existing data-driven methods caused by data quality problems and the frequent neglect of expert prior knowledge, a data-knowledge fusion risk assessment method for urban water supply networks was proposed based on GPC and GMM. First, risk scores were obtained by expert scoring rules and were used as the basis for knowledge fusion. Second, a feature matrix was constructed by combining the scores with objective attributes, including pipeline physical properties and operating environment. A GPC model was used to predict the failure probability of each pipe segment. Third, cluster analysis was performed on the predicted probability distribution, and a Gaussian mixture model was applied to classify risk levels scientifically. Finally, experiments were conducted on a dataset of pipe segments and failure records from Kitchener, Canada. The results show that the proposed data-knowledge fusion model performs well under different data scenarios, and at a 1% replacement rate, up to 30% of potential failures are prevented. The area under the curve of 0.94 is achieved for risk ranking. The chi-square coefficient for risk classification exceeds 0.1, and the coefficient of determination is greater than 0.7. SHAP analysis verifies the key role of expert scoring.

Key words: data-knowledge fusion, urban water supply network, risk assessment, gaussian process classification (GPC), Gaussian mixture model (GMM), SHapley additive explanations(SHAP)