As part of an optimized maintenance strategy, track monitoring should provide information to predict track faults at an early stage. This is possible by continuously measuring the axle box accelerations and using artificial intelligence, which can detect short wave defects on the railway track with high accuracy. Such short wave defects include rail breaks, cracks, and local irregularities (mud spots). These types of faults can reduce the track quality in a short period of time.
Different track irregularities were simulated in a track-vehicle scale model to generate acceleration data for typical track defects. The main focus of the current research is on recognition of local irregularities in the track-vehicle scale model. To implement the artificial intelligence, a Recurrent Neural Network is used to show the procedure and the results of recognition of track defects. The architecture and components of the neural network used are described in detail in this article. At the end of the article, a table summarizing the results of the different models trained for detecting the local irregularities in the track-vehicle scale model is presented.
artificial intelligence, deep learning, detection, local instability, maintenance, railway
 Rapp, S. Martin, U. Strähle, M. & Scheffbuch, M., Track-vehicle scale model for evaluatinglocal track defects detection methods. Transportation Geotechnics, 19, pp. 9–19,2019. https://doi.org/10.1016/j.trgeo.2019.01.001
 Bahamon-Blanco, S. Rapp, S. Rupp, C. Liu, J. & Martin, U., Recognition of trackdefects through measured acceleration P1 & P2. 7th International Conference ofEACEF (European Asian Civil Engineering Forum), 615, 2019.
 Rapp, S. & Martin, U., Erkennung von punktuellen Unstetigkeitsstellen am Fahrwegam Beispiel eines Fahrzeug-Fahrwegmodells – Ansatz zur Modellbildung (EPIB 1.1).Auftaktworkshop EPIB, 2018.
 Bahamon, S., Detection of local instabilities on a scale vehicle-track model throughmeasured accelerations of the vehicle. Master Thesis, Institute of Railway and TransportationEngineering of the University of Stuttgart, 2018.
 Sander, K., Anwendung Beschleunigungssensormodul Version 3.Kurzform, 2019.
 Zaccone, G., Karim, R. & Menshawy, A., Deep Learning With TensorFlow: ExploreNeural Networks With Python. Packt Publishing, p. 72, 2017.
 Aggarwal, C., Neural networks and deep learning. Springer international publishingAG, p. 38, 2018.
 Deep Learning: Long Short-Term Memory Networks (LSTMs), Online. https://www.youtube.com/watch?v=5dMXyiWddYs. Accessed on: 4 May 2020.
 Hopkins, B., A Wavelet-Based Rail Surface Defect Prediction and Detection Algorithm.Virginia Polytechnic Institute and State University, pp. 74–82, 2012.
 How to Reduce Overfitting With Dropout Regularization in Keras, Online. https://machinelearningmastery.com/how-to-reduce-overfitting-with-dropout-regularizationin-keras. Accessed on: 10 May. 2020.
 Zhang, Y., Work Summary Deep Learning. Institute of Railway and TransportationEngineering of the University of Stuttgart, 2019.
 Goodfellow, I., Deep Learning. mitp Verlags GmbH & Co, p. 6, 2018.
 NN-SVG Publication-ready NN-architecture schematics, Online. http://alexlenail.me/NN-SVG/index.html. Accessed on: 4 May 2020.
 Yang, B., Class of Deep learning Chapter 5: Advanced optimizaion thechniques, Instituteof signal processing and system theory. of the University of Stuttgart, 2018.
 Srivastava, N., Dropout: A Simple Way to Prevent Neural Networks from Overfitting.Journal of Machine Learning Research, 15, pp. 1929–1958, 2014.
 What is the role of the activation function in a neural network? How does this functionin a human neural network system? Online. https://www.quora.com/What-is-the-roleof-the-activation-function-in-a-neural-network-How-does-this-function-in-a-humanneural-network-system2018. Accessed on: 4 May 2020.