Estimating the Values of Missing Data Related to Infrastructure Condition States Using Their Spatial Correlation

Abstract

Infrastructure managers consistently monitor the condition of their assets to predict their deterioration speed and determine the optimal time to execute preventive interventions. However, despite the recent progress in more frequent and accurate monitoring of assets and storage of the related results, in practice, real-world data often contains errors and discrepancies such as missing data or faulty entries. This problem can happen owing to collection errors during routine inspections or inconsistency of data storage formats in different years. Because the quality of data plays a significant role in the accuracy of deterioration prediction and the resulting intervention programs, it is important to improve condition state predictions by imputing the values of missing information. This paper examines the efficiency of different models that use the spatial correlation of infrastructure assets in predicting the value of missing data. The models include univariate and multivariate Kriging, a hybrid artificial neural network (ANN)-Kriging model, and the bidirectional long short term memory (bi-LSTM) neural network, which can model the data with spatial correlation or a sequential relationship. The results confirm that the condition indicator values can be estimated with reasonably low levels of error.

A Probabilistic Approach for Missing Data Imputation

“…. ANN [29] is a type of machine learning algorithm inspired by the structure and function of the human brain. It consists of a network of artifcial neurons that are interconnected to process information and make predictions.…”