Predicting atmospheric corrosion of low-alloy steels based on environmental features extracted by improved autoencoder

Material corrosion can cause substantial economic losses and pose serious damages. One way to reduce those damages is to predict available data before catastrophic occurrences. However, because of the small sample size of the corrosion data, the generalization of the corrosion prediction models remains a challenge. This study adopts an autoencoder to determine the intrinsic features of the natural atmospheric environment to which the materials were exposed. The training process of the autoencoder was designed to ensure that the features were consistent between natural atmospheric environmental data and the averaged environmental data within small-sample corrosion data. In addition, the k-nearest method was used to find a set of extracted features for the corrosion rate data. Thus, the amount of corrosion data increased. In this study, a neural network was adopted for augmented corrosion data to predict corrosion rates. The experimental results showed that the proposed method not only had a satisfactory training accuracy but also a strong generalization ability.