@article{3416,
  author = {An Approach Based on Time-Series and Neural Networks for Safety Railway  Incident Prediction},
  title = {An Approach Based on Time-Series and Neural Networks for Safety Railway  Incident Prediction},
  journal = {Digital Signal Processing and Artificial Intelligence for Automatic Learning},
  year = {2022},
  volume = {1},
  number = {1},
  doi = {https://doi.org/10.6025/dspaial/2022/1/1/11-27},
  url = {https://www.dline.info/dspai/fulltext/v1n1/v1n1_2.pdf},
  abstract = {Every day, thousands of people travel by train, not escaping the leniency of unforeseen events, the fluidity of the rail network being able to be disrupted by equipment breakdowns. Therefore, predictive maintenance is relevant and necessary to help anticipate these breakdowns and thus act against any mechanical, electrical, or technical constraints or obstacles that could disrupt or prevent the normal circulation of trains. This process is carried out using artificial intelligence approaches and various machine learning and deep learning models. This article will implement an approach that combines two essential concepts: time series and neural networks. We will start with the univariate analysis of the number of failures per day using a range of machine learning and deep learning algorithms, namely LSTM, BiLSTM, GRU, and SVR. The results show that we manage to minimize the prediction error; for example, with the GRU model, we get an RMSE of 0.487, but with increasing data, we get an RMSE of 0.463.

Moreover, the problem encountered is the detection of peaks; the models cannot detect outliers, hence the use of the SVR model, which gives better coordination between the test data and the predicted data, with a gamma value of 0.03. Then, we tested the VAR model and the LSTM with several outputs; the latter gives satisfactory results with an accuracy rate of 92% and an RMSE of 0.006. Finally, we address the problem of classification of the nature of failures. We used several machine learning algorithms, such as SVM, KNN, Random Forest, then tested a method of “Ensemble Learning,” the Vote. In the latter, we combined the three algorithms used previously, which increased the accuracy of the test to 61.73%. },
}