Earthquake Damage Prediction of Underground Steel Pipe with Screw Joint Using Machine Learning

In this study, damage to old type steel conduits is predicted using machine learning. We used a supervised machine learning algorithm with a gradient boosting method. To construct the estimation model, we use inspection data from five past large earthquakes: the 1995 Southern Hyogo Prefecture Earthquake, Mid Niigata Prefecture Earthquake in 2004, the Niigataken Chuetsu-oki Earthquake in 2007, the 2011 off the Pacific coast of Tohoku Earthquake, and the 2016 Kumamoto Earthquake. The damage was checked by pipe inspection camera. The model was constructed using 75% of the 23,833 records of the inspection results, and a test was carried out using the remaining 25% of the records. Each record has a label (damaged or no damage), and has 14 variables. Using this method, we constructed a model which has high prediction accuracy, an area under the receiver operating characteristic curve of 0.87 with the test data. The converted displacement in this model contributed the most by confirming the contribution of the respective variables. Old steel conduits with screw joint cannot absorb displacement, and the estimation model may be interpretable.