Prédiction & classification des Incidents Ferroviaires du Matériel Roulant.

Abstract #

The rail transport system is made up of industrial assets that require considerable investment in

terms of maintenance. This applies to rolling stock such as trains, locomotives, wagons, and

even more to infrastructure such as tracks and tunnels. However, environmental conditions or

mechanical forces can lead to degradation of the tracks and machines.

In this context, the objective of this report is to design and implement a predictive solution of

future incidents for the Office National des Chemins de Fer.

This predictive analysis uses historical data of the ONCF, in the form of time series, this

information is stored to allow the realization of several analyses and to make more informed

strategic decisions for the future.

After a good documentation of the different techniques dedicated to solve the problem, namely,

the machine learning algorithms and those of deep learning, then, we conceptualized our project

with a process to follow during the realization of this project, to finally start the realization part

which includes three main parts, the prediction of the number of breakdowns per day, the

prediction of the delays of the breakdowns (direct delay, indirect delay and cumulative delay)

and the classification of the nature of the railway breakdowns.

First, for the prediction of the number of failures per day, we tested a range of models, namely,

ARIMA, LSTM, BiLSTM, GRU and SVR, the latter with better results in terms of coordination

of test and predicted data. Then, for the prediction of the failure delays, we used the VAR model

and LSTM, after the evaluation, we could conclude that LSTM gives better results of accuracy

with 92% and an RMSE equal to 0.006. Finally, for the classification of the nature of failures,

we tried several Machine Learning models, such as k-Nearest Neighbours, SVM, Random

Forest…

Keywords: Prediction, Railway incidents, Data Science, Time Series, Machine learning, Deep

learning, ARIMA, LSTM, SVM, GRU.