12th International Conference on Zinc & Zinc Alloy Coated Steel Sheet - GALVATECH 2021

For prediction of corrosivity of steel products used in atmospheric environments, exposure tests in the actual environment have been widely conducted. Long-term corrosion loss has been predicted by extrapolating the change of corrosion loss with time in exposure tests. However, the data of exposure test in regions where we would like to know corrosion rate doesn’t always exist. In this case, exposure test has to be newly started, and it consumes much time to obtain long-term data. For this reason, some atmospheric corrosion prediction formula which is composed of relation between corrosion loss and environmental factors, like temperature and amount of airborne sea salt, have been suggested. One example is dose-response function defined in ISO9223. However, it is difficult to predict long-term corroison loss with high aquracy by using the formula, since there is limitation on building one formula from complex relation between corrosion loss and environmental factors. In contrast to that, data science is considered to be useful for anlayzing complex data. In this study, data science was applied to atmospheric corrosion of carbon steel. The most suitable prediction model of data science was selected on the basis of experimental data analysis, and new prediction method for atmospheric corrosion was suggested. To increase accuracy, we established a prediction model which can treat the following three characteristics of exposure test data. (1) Multicollinearity: there is correlation between environmental factors. (2) Nonlinearity: there is nonlinear relation between corrosion loss and environmental factors. (3) Small data sets: normally, large data sets are needed to increase prediction accuracy, however, we have small data sets since exposure test takes a long time and many works are involved. In our model, we chose temperature, relative humidity, amount of airborne salt and SO2 as input variables to predict the corrosion loss. To solve multicollinearity, our model uses latent variables as explanatory variables instead of input varialbes. The latent variables are created by transforming the input variables orthogonally and can explain the variance direction in the response variable. In addition, by using the locally weighted regression, "nonlinearity" was treated. It predicts by a linear regression model fitted strongly to the training data that are closer to the data we are trying to predict. Thus, high prediction accuracy is secured even if small data sets. This method gave us an accurate prediction of corrosion loss in comparison with dose-response function in ISO9223.