The RAIL-D project was created with the goal of transforming railway maintenance, shifting from a reactive to a predictive approach. By integrating Machine Learning and Deep Learning algorithms, combined with Data Augmentation techniques and physics-informed constraints, a system has been developed that can analyze multi-sensor data from diagnostic trains and predict potential degradation areas in advance.The platform leverages an ensemble of models (Random Forest, XGBoost, and LSTM) which, working in synergy, achieved an accuracy close to 90% in fault prediction, significantly reducing false alarms and improving intervention planning. An innovative aspect is the use of synthetic data generated with realistic engineering constraints, which has enriched the training datasets and enhanced the robustness of the system. Furthermore, the development of digital twins allows simulation of infrastructure evolution scenarios, providing a strategic support tool for managers and operators. The results obtained demonstrate a tangible impact: fault anticipation up to two months in advance, reduced maintenance costs, and improved service reliability. RAIL-D therefore represents an important step towards a safer, more efficient, and digital railway, contributing to the sustainable mobility of the future.
case-studies
RailD – Rail Detection
Approach and methodology
The RAIL-D project was designed to redefine the paradigm of railway maintenance, moving beyond the traditional fault-based approach and introducing next-generation predictive tools. At the core of the initiative lies the integration of Machine Learning and Deep Learning models with Data Augmentation techniques and physics-informed principles, which allow the fusion of data analysis with the real physical constraints and behaviors of infrastructures. In this way, the information flows collected by diagnostic trains are not merely recorded but transformed into dynamic and contextualized insights. The platform is based on an ensemble of heterogeneous algorithms – such as Random Forest, XGBoost, and LSTM – that operate in synergy, leveraging their respective strengths to make predictions more reliable. The use of synthetic data generated through realistic engineering constraints also makes it possible to expand the available datasets, strengthening the system’s ability to adapt to complex scenarios and rare situations. Added to this is the development of digital twins, true digital replicas of the infrastructure, which make it possible to simulate different evolutions over time and test maintenance strategies in a virtual environment before applying them in the field.
RAIL-D adopts an integrated approach that combines Machine and Deep Learning (Random Forest, XGBoost, LSTM) with multi-sensor data and physics-informed techniques, enhanced by the use of digital twins. Thanks to augmented data and simulations, the system ensures more accurate, robust, and reliable predictions in support of predictive maintenance.
Staff involved in the project
4
Specialists
Execution time
10
Months
Technologies used
- AWS Bedrock
- AWS Forecast
- Database vettoriale
- Python
50
%
Prediction
-90
%
Fault prediction accuracy
2
Average months in advance for fault prediction
Get in touch with
Abacus Group
Contact Us
Thank you for providing the required information below, marked with an asterisk (*). This will help us provide you with the best possible response.