Signal separation and continuous missing value imputation of strain gauge in the icebreaker sensor system

Abstract

Korea's first icebreaking research vessel, ARAON, has been conducting icebreaking performance tests in polar waters since 2010. These tests collect data to calculate design factors for the structural safety and load optimization of icebreakers. However, due to ARAON's operation in extreme environments, missing values frequently occur in the collected data, leading to unreliable results. This study focuses on imputing missing values from the three-axis strain gauges among various sensors attached to ARAON. Strain gauges, located on the hull plate where collisions with ice occur for local ice load estimation, experience a high frequency of missing values. Imputing missing values from strain gauges presents two major challenges. The first challenge is handling the composite signal. Strain data is a mixture of response signals caused by various external factors, such as hull-ice collisions and engine noise, making it difficult to impute missing values using raw data alone. The second challenge is with continuous missing data. There are still limitations to imputing continuous missing data, which is commonly encountered in real-world scenarios. To address these challenges, locally weighted scatterplot smoothing (LOWESS) regression and Tukey's fences method are used to separate the strain data according to strain components. Subsequently, the within and inter-sensor relationship-based linear regression (WIRLI) is proposed to have good imputation performance even for continuous missing data. The results demonstrate that the WIRLI performs robust imputation even when the missing rate changes. Hence, this model can be applied to impute the missing values in icebreaking performance tests where various missing scenarios occur. ？ 2024