Toward smart water infrastructure monitoring: Bayesian-optimized interpretable deep learning for pipeline failure prediction

Purpose Deep learning approaches are commonly employed to develop predictive maintenance solutions utilizing large datasets. However, their black-box framework, performance limitations for tabular big data and demand for high-fidelity data limit their adoption. Thus, the study proposes an interpretative deep learning framework with a transformer structure to support tabular data. Design/methodology/approach The Bayesian optimized tabular interpretative deep learning (β-TabNet) framework comprises a multi-layer transformer architecture. The models utilize an attentive transformer for stage-wise feature selection and a linear mapping layer for predicting time-to-next failure. A Bayesian optimization module is then used for cross-validation to tune the model parameters and minimize average error values. Developed models are tested on unseen data and compared with other published approaches for validation. Findings Utilizing the β-TabNet model, the research provides interpretable, high-accuracy failure predictions with mean error ranges between 0.67 and 2.47, correlation coefficient values above 0.98, and root mean square values of ∼0.9. The base models are also benchmarked against the performances of other published models. Models highlight the installation year and age at last failure as the strongest predictors. Furthermore, there is a strong interlinkage between pipeline failure and ground conditions, with a decrease in time-to-next failure with an increase in soil corrosivity and urbanization. Social implications Accurate water leak detection reduces water waste and prevents costly property damage, promoting more equitable access to affordable water utilities. It also minimizes disruptive infrastructure repairs and service outages, enhancing community resilience and public trust in municipal systems. Originality/value The knowledge contribution of this article is beneficial to water reticulation and asset management for city networks and trace urbanization impacts on increasing pipeline failure. Using the model, the authorities can foresight and reduce downtime of water infrastructure by predicting through automated computations.