Periodic Reporting for period 1 - RAILS (Roadmaps for A.I. integration in the raiL Sector)
Periodo di rendicontazione: 2019-12-01 al 2021-09-30
The Shift2Rail project RAILS addresses the topic “Artificial Intelligence for the railway sector”. The impact of Artificial Intelligence (AI) techniques on the industry has been so disruptive that it gave rise to a new wave of research and applications. Several industries, mainly in automotive, robotics, logistics, and manufacturing, have benefited significantly from AI adoption and this positive trend is planned to be kept in the future. Modern and future railways represent one of the fields in which AI is expected to have a significant impact from a medium to long-term perspective. This raises many issues regarding system safety, dependability, security, and new threats associated with a high level of autonomy, but it also opens for unprecedented scenarios and great opportunities.
The overall objective of the RAILS research project is to investigate the potential of AI in the rail sector and contribute to the definition of roadmaps for future research in next-generation signalling systems, maintenance, and network management.
The main objectives of the project are as follows:
- Objective 1: Identification of the potential of AI for railways;
- Objective 2: Adherence to current work in railways innovation;
- Objective 3: Recognition of required innovation shifts;
- Objective 4: Development of methodological and experimental proofs-of-concept;
- Objective 5: Development of Benchmarks, Models, and Simulations;
- Objective 6: Transition pathways toward the rail system scenario;
- Objective 7: Involvement of relevant rail stakeholders;
- Objective 8: Training of young researchers and creation of a research network on AI in railways.
The overall objective of the RAILS research project is to investigate the potential of AI in the rail sector and contribute to the definition of roadmaps for future research in next-generation signalling systems, maintenance, and network management.
The main objectives of the project are as follows:
- Objective 1: Identification of the potential of AI for railways;
- Objective 2: Adherence to current work in railways innovation;
- Objective 3: Recognition of required innovation shifts;
- Objective 4: Development of methodological and experimental proofs-of-concept;
- Objective 5: Development of Benchmarks, Models, and Simulations;
- Objective 6: Transition pathways toward the rail system scenario;
- Objective 7: Involvement of relevant rail stakeholders;
- Objective 8: Training of young researchers and creation of a research network on AI in railways.
AI is a very wide research field and the term AI is often improperly used to address closely related concepts or applications that can benefit from AI but cannot be considered AI techniques. To provide a baseline for research activities in railway transportation, a Taxonomy has been proposed that classifies AI-related concepts and describes their relationships using a semi-formal notation.
Then, a comprehensive review of the current research of AI in railways has been conducted to investigate the relevant state-of-the-art of AI approaches and identify: i) the most addressed areas as well as the railway areas that seem not have been researched yet; ii) the current mapping between AI techniques and railway problems and the most interesting research directions. This review includes both the scientific literature and all the Shist2Rail projects, past or ongoing. Further European projects, as well as some overseas projects, have been considered, too.
Figure 1 shows the railway domains addressed by the project, ordered by the number of scientific papers dealing with the application of AI techniques in those domains. The analysis reveals that the most researched domain for AI application is maintenance and inspection, on the other extreme, almost no research work investigates the application of AI techniques in Revenue Management, followed by Transport Policy and Autonomous Driving and Control. The pie chart in Figure 2 shows that the distribution of the projects addressing AI in the same railway domains confirms this result.
The analysis was extended through a Survey on Challenges and State-of-Practice of AI in the railway sector. A questionnaire was submitted to the railway stakeholders to identify the main practical issues that have to be overcome to effectively apply AI in railway applications, the most blocking factors, and the key milestones to be reached. The findings of the Survey, the state-of-the-art, and the contribution of the Advisory Board members allowed depicting a comprehensive picture of the relevant railway problems that the railway stakeholders would tackle and the AI techniques currently adopted, they are reported in Figure 3.
The collected information allowed prioritizing the application areas for AI in railways. Urgent, High Impact and Promising areas have been suggested. The urgent class requires addressing the problem of Data Availability, the relationship between Safety and Trustworthy AI, the need for Standards and Regulations, and investigating the potential of AI for Risk Assessment. The High Impact class includes railway applications that could receive great benefits from an effective application of AI due to the high maturity level already reached in the research fields, such as in Maintenance, or because of the high positive impact they could have on the system safety, on the level of automation of Automatic Train Operation, or in Traffic Planning and management; finally, the Promising class encompasses new or uncharted directions, such as Revenue Management, Passenger Mobility, Traffic Signalling and Dynamic Route Selection.
Among the ways for a fast take-up of AI in railways, the transferability of results and approaches from other sectors is a key issue. A review is ongoing of AI applications currently adopted or investigated in sectors such as automotive, robotics, maritime, avionics, and the like. In order to provide a framework for transferability studies a qualitative approach has been proposed that will be further developed and a set of case studies that will provide the context for proof of concepts have been identified. The set of benchmark case studies includes Railway Obstacle Detection and Collision Avoidance for Intelligent Train Operation, Cooperative Driving for Virtual Coupling of Autonomous Trains, Remaining Useful Life evaluation, Prognostic technology for railway maintenance.
RAILS addresses the training of two Ph.D. students to support the research capacity in AI within the rail sector across Europe by involving research institutions with a combined background in both computer science and transportation systems. To support the creation of a research network on AI in railways a specific series of workshops, AI4RAILS, has been settled whose first and second editions have been held in 2020 and 2021.
Figure 4 provides a bird's eye view of the project in the first period.
Then, a comprehensive review of the current research of AI in railways has been conducted to investigate the relevant state-of-the-art of AI approaches and identify: i) the most addressed areas as well as the railway areas that seem not have been researched yet; ii) the current mapping between AI techniques and railway problems and the most interesting research directions. This review includes both the scientific literature and all the Shist2Rail projects, past or ongoing. Further European projects, as well as some overseas projects, have been considered, too.
Figure 1 shows the railway domains addressed by the project, ordered by the number of scientific papers dealing with the application of AI techniques in those domains. The analysis reveals that the most researched domain for AI application is maintenance and inspection, on the other extreme, almost no research work investigates the application of AI techniques in Revenue Management, followed by Transport Policy and Autonomous Driving and Control. The pie chart in Figure 2 shows that the distribution of the projects addressing AI in the same railway domains confirms this result.
The analysis was extended through a Survey on Challenges and State-of-Practice of AI in the railway sector. A questionnaire was submitted to the railway stakeholders to identify the main practical issues that have to be overcome to effectively apply AI in railway applications, the most blocking factors, and the key milestones to be reached. The findings of the Survey, the state-of-the-art, and the contribution of the Advisory Board members allowed depicting a comprehensive picture of the relevant railway problems that the railway stakeholders would tackle and the AI techniques currently adopted, they are reported in Figure 3.
The collected information allowed prioritizing the application areas for AI in railways. Urgent, High Impact and Promising areas have been suggested. The urgent class requires addressing the problem of Data Availability, the relationship between Safety and Trustworthy AI, the need for Standards and Regulations, and investigating the potential of AI for Risk Assessment. The High Impact class includes railway applications that could receive great benefits from an effective application of AI due to the high maturity level already reached in the research fields, such as in Maintenance, or because of the high positive impact they could have on the system safety, on the level of automation of Automatic Train Operation, or in Traffic Planning and management; finally, the Promising class encompasses new or uncharted directions, such as Revenue Management, Passenger Mobility, Traffic Signalling and Dynamic Route Selection.
Among the ways for a fast take-up of AI in railways, the transferability of results and approaches from other sectors is a key issue. A review is ongoing of AI applications currently adopted or investigated in sectors such as automotive, robotics, maritime, avionics, and the like. In order to provide a framework for transferability studies a qualitative approach has been proposed that will be further developed and a set of case studies that will provide the context for proof of concepts have been identified. The set of benchmark case studies includes Railway Obstacle Detection and Collision Avoidance for Intelligent Train Operation, Cooperative Driving for Virtual Coupling of Autonomous Trains, Remaining Useful Life evaluation, Prognostic technology for railway maintenance.
RAILS addresses the training of two Ph.D. students to support the research capacity in AI within the rail sector across Europe by involving research institutions with a combined background in both computer science and transportation systems. To support the creation of a research network on AI in railways a specific series of workshops, AI4RAILS, has been settled whose first and second editions have been held in 2020 and 2021.
Figure 4 provides a bird's eye view of the project in the first period.
The ongoing road-mapping activities will effectively contribute to defining the future research directions and the design and implementation of smarter railways. Some steps in this line are provided by the findings of the first phase of the project as described above. Figures 5 and 6 list the main obstacles and milestones according to the opinion expressed by the railway stakeholders who participated in the Survey. With respect to these issues, some preliminary recommendations emerge from the work done so far pointing at:
1. The appointment of working groups to define pilot case studies/demonstrators and challenges for the academic and industry community with the aim of producing the knowledge needed to drive standardization.
2. The funding of specific actions for railway dataset generation, sharing, processing, and management. This may also include the definition of alliances/federations among relevant stakeholders to share data.
3. The promotion of data-driven approaches aiming at enhancing the safety of passengers and workers, such as data-driven risk assessment, accident prediction, avoidance, and analysis.
4. The support for the development of assistant AI applications for safe functions, in particular exploiting the concept of safety envelops to decouple perception and safety-critical tasks.
1. The appointment of working groups to define pilot case studies/demonstrators and challenges for the academic and industry community with the aim of producing the knowledge needed to drive standardization.
2. The funding of specific actions for railway dataset generation, sharing, processing, and management. This may also include the definition of alliances/federations among relevant stakeholders to share data.
3. The promotion of data-driven approaches aiming at enhancing the safety of passengers and workers, such as data-driven risk assessment, accident prediction, avoidance, and analysis.
4. The support for the development of assistant AI applications for safe functions, in particular exploiting the concept of safety envelops to decouple perception and safety-critical tasks.