PROPOSAL OF AN EMERGENCY PREDICTION MODEL FOR SEWER PIPE CULVERTS USING NEURAL NETWORK -COMPARISON OF ACCURACY BY CHANGING HYPERPARAMETERS-

Abstract

The total length of sewer culverts in Japan is currently about 490,000 km, of which about 25,000 km have a standard service life of 50 years, and this number is expected to increase rapidly in the future. However, the current inspection methods of visual inspection and television cameras require an enormous amount of time to inspect all sewer culverts. To solve this problem, it is necessary to determine the necessity and priority of inspection. In this study, we constructed models to predict the deterioration status of sewer culverts using only specification data and surrounding environmental data, without the need to survey the inside of the culvert to obtain ANN data, using sewer pipe data made available by the National Institute for Land Infrastructure Management of Sewerage Research. Several models with different hyperparameters (e.g., optimization algorithm) and number of hidden layers were created, and their classification performance was compared to determine the appropriate hyperparameters.