Deep Leaning Neural Networks for Determining Replacement Timing of Steel Water Transmission Pipes

Water main pipe breaks are an ongoing concern worldwide. Large-diameter steel water transmission mains (WTMs) transport a much larger volume of water and their failure leads to even greater damages than those seen in water networks with small diameter iron or PVC pipe lines. However, there is no predictive model for large-diameter steel WTMs, leaving retroactive maintenance as the sole means of prevention. The objective of this study was to predict the optimal replacement timing for large-diameter steel WTMs based on physical and environmental factors, using Deep Learning algorithms. The model was developed in four steps: (1) determine major factors, (2) determine the best model by comparing performances of three neural networks (NNs) (a shallow artificial NN, multiple hidden layered NN, Stacked autoencoder NN), (3) classify the data into homogeneous groups by an ANN-based clustering technique, and (4) perform the developed model for each group. The multiple hidden layered NN was found to be the best deep neural NN in forecasting a replacement timing of aging WTMs. Additionally, it is recommended that such ANN-based clustering methods be used in predicting a more accurate replacement timing of water networks and making a quantitative decision on replacement.