CORROSION PREDICTION METHOD FOR INNER SURFACE OF TELECOMMUNICATION CONDUIT WITH MACHINE LEARNING BASED ON INSPECTION RESULTS

Abstract

In order to predict earthquake damage to underground telecommunication pipelines in advance, it is necessary to predict corrosion on the inner surface of the pipelines. We developed a prediction model using gradient boosting, a type of machine learning, to determine the presence or absence of corrosion. The variables used in the prediction were equipment data such as elapsed years, and installation environment data such as elevation, climate, watershed, and soil. A prediction model with an ROCAUC of 0.86 and an f-value of approximately 0.56 could be created for the evaluation data, confirming its versatility. Analysis of the model showed that it predicts the corrosion rate based on the elapsed time and the corrosion rate of each river basin, and these variables are considered to be effective. Discarding variavles that are difficult to interpret such as length and modification of seawater areas may improve the model performance.