Estimating the Values of Missing Data in Railway Networks Using their Spatial Correlation

Abstract: In railway infrastructure management, it is always desired to know the conditionof each track segment, which is represented approximately by the conditionindicators. The values of condition indicators can be used as a decisioncriterion about whether the intervention on the track should be implementedin the future or when the intervention should be executed. However, the informationis not sufficient to draw a decision or a prediction when working withthe condition indicator values. This problem can happen due to collection errorswhen inspecting the track or the inconsistency of the data storage formatin different years or in different notations. Therefore, fulfilling the competenceof the track segment condition data sets by estimating the missing data tosupport the decision-making process is important. In this paper, experimentsof different models that can utilize the spatial correlation among data pointsare done to investigate their estimation ability. The models include the Kriging,Co-Kriging, ANN-Kriging hybrid model and Bi-LSTM neural network,which are all having the ability to model the data with spatial correlation or asequential relationship. The condition indicator values of the track segmentsare used and serval auxiliary variables correlated to the condition values arealso included. The results show that the condition values could be estimatedwith reasonably low estimation errors based on their potential correlation.