Handling Incomplete and Missing Data in Corrosion Pit Measurement Database Using Imputation Methods: Model Development Using Artificial Neural Network

摘要

Data scarcity and missing values are a prime challenge in developing a corrosion prediction model. In this paper, eight imputation techniques are explored using the National Bureau of Standards (NBS) corrosion database. The eight imputation techniques are mean, median, linear regression (LR), K-nearest neighbor (KNN), iterative robust model-based imputation (IRMI), multiple imputations of incomplete multivariate data (AMELIA), sequential imputation for missing values (IMPSEQ), and principal component analysis (PCA). The utility of imputation techniques is checked by training a neural network (NN) on the data sets imputed by the eight techniques. Among the techniques, KNN and IMPSEQ performed better by achieving a low error and high coefficient of determination R2. Results were compared with a baseline accuracy, where the NN model was trained on the original corrosion data set without the missing values. The NN performance increased from the baseline accuracy (81%) when it was trained by KNN (85%) and IMPSEQ (91%) imputed data sets.