Abstract

The prediction of random phenomena has long seemed impossible to achieve without the arrival of Machine Learning techniques and the development of computer power. These techniques have been applied in several fields but not in rail transport, except for some work on the presence of faults on rails and machines. This work starts from the results obtained by the k-nearest neighbors algorithm (k-NN) method used previously (87% with the ROC curve and 83.61% with the confusion matrix for 3 neighbors) to show that it is possible to improve this ratio. This work proposes to observe the possibilities of considering a classifier resulting from the combination of existing classifiers to produce another which could give a higher ratio because human lives depend on it. The classifier to be implemented will consider the condition of the equipment, the loading, the drivers, and especially the condition of the track. Predicted in this way, vehicles predicted to derail can be removed from the train and repaired, then resubmitted to the predictor. The train can only be authorized to depart if the number of vehicles to be derailed drops to zero. Such a prediction will surely save human lives and materials and make rail transport more reliable

Key words: Machine learning, derailment, prediction, k-nearest neighbors algorithm (k-NN), combination of classifiers, fuzzy data, neuro-fuzzy classifier, risk.