• predictive maintenance
• machine learning
• railway
• data-driven

The use of big data has contaminated industries 4.0 revolutionizing their maintenance strategies.
Railway industry, due to the extensive use of locomotives and infrastructure and the incremental focus on user service, has patented data collection systems so that their analysis allows to consolidate the decision-making process aimed at optimizing resources and minimizing service disruption.
The migration from condition-based maintenance to a predictive approach is sometimes hindered and slowed down by legislative limits that prescribe the transparency of processes, making the application of black box algorithms laborious for reasons of explainability.
In order to analyze the feasibility of a scalable maintenance solution based on machine learning algorithms, this thesis extracts the railway database of Italian and Swiss high-speed train runs in 2019 by reverse engineering the condition-based software of the company object of the case study.
The training of a rail classifier able to discriminate if a locomotive needs or not a maintenance action was carried out by means of Decision Tree for which it emerged the criticality of complex interpretability caused by the high number of features as well as the risk of overfitting.
It was then proposed the modeling with Random Forest which through the study of the features importance is proposed to select the most significant variables for the analysis with respect to the target variable.
Finally, the time series of the occurrences of the daily aggregated alarms were considered in order to predict the trend and facilitate the dynamic programming of the workload necessary for the remote management of maintenance assistance.
The serie was decomposed in order to separate the systematic component from the random one and, after the exponential smoothing process, the autoregressive integrated moving average model was applied to make the forecast, which turned out to be statistically significant.