Keywords: railway track degradation, deep learning, track geometry index, track recording machine, EM120.

Rail track deterioration models are integral components of rail infrastructure maintenance management systems. In particular, track geometry defects are one of the leading causes of train accidents. Also, control, management, and modification of geometric conditions are one of the most important tasks of railway maintenance management systems. Track geometry data such as profile, alignment, gauge, cross-level, and twist constantly change over time. Therefore, these features have the characteristics of time series data. In this study, a large database from outputs of EM120, a track recording machine, was provided for the years 2009 to 2020 and for all 19 railway zones of Iranian Railways (approximately 14,000 km of railway track and 100 GB of data). From Deep Learning techniques, CNN, LSTM, and CNN-LSTM models were selected to predict track geometry degradation. Long short-term memory (LSTM) has the advantage of analysing relationships among time-series data through its memory function, while CNN models may filter out the noise of the input data and extract more valuable features that would be more useful for the final prediction model. By integrating convolutional neural networks (CNN) with long short-term memory (LSTM), a CNN-LSTM model is considered to be more accurate and can make better point-wise predictions. The models were built from the average segments of 100 and 200 meters. The forecasting results of proposed models were analysed and compared, and the CNN-LSTM model with a segment length of 200 m and sequence length of 6 reported the best forecasting performance, achieving an R-squared value of 0.913.

download the full-text of this paper (PDF, 6 pages, 232 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description