Data of sufficient quality, quantity and validity constitute a sometimes overlooked basis for eMaintenance. Missing data, heterogeneous data types, calibration problems, or non-standard distributions are common issues of operation and maintenance data. Railway track geometry data used for maintenance planning exhibit all the above issues. They also have unique features stemming from their collection by measurement cars running along the railway network. As the track is a linear asset, measured geometry data need to be precisely located to be useful. However, since the sensors on the measurement car are moving along the track, the observations’ geographical sampling positions come with uncertainty. Another issue is that different seasons and othertime restrictions (e.g. related to the timetable) prohibit regular sampling. Hence, prognostics related to remaining useful life (RUL) are challenging since most forecasting methods require a fixed sampling frequency.

This paper discusses methods for data cleaning, data condensation and data extraction from large datasets collected by measurement cars. We discuss missing data replacement, dealing with autocorrelation or cross-correlation, and consequences of not fulfilling methodological pre-conditions such as estimating probabilities of failures using data that do not follow the assumed distributions or data that are dependent. We also discuss outlier detection, dealing with data coming from multiple distributions, of unknown calibrations and other issues seen in railway track geometry data. We also discuss the consequences of not addressing or mishandling quality issues of such data.