A hybrid deep survival model for failure modeling of water distribution networks coupling physical survival and data reconstruction

Abstract

To detect and replace damaged pipes and maintain the stable operation of water supply systems timely, it is crucial to carry out pipeline failure prediction. Limited by the feature nonlinear ability and generalization ability of survival machine learning, the application effect in pipeline failure prediction is not always satisfactory. To develop more efficient, powerful, and flexible prediction models, a hybrid deep survival model (HDSM) is proposed by coupling deep auto-encoder and survival analysis to effectively predict pipeline failures. Guided by the theory of mechanism and data fusion, the data reconstruction constraints and survival analysis constraints are coupled into the loss function by HDSM. Its dual advantages of combining the powerful feature extraction capability of deep auto-encoder and the statistical inference role of survival analysis can more accurately predict the survival time of physical pipeline systems. In a real pipeline network, the superiority and effectiveness of HDSM are verified in comparison with other deep survival learning and survival machine learning, with C-index exceeding 0.95 and Brier score below 0.056. Finally, sensitivity analyses of different hyperparameters are carried out to verify the robustness of the HDSM model in pipeline failure prediction.