As the railway industry advances, integrating Artificial Intelligence (AI) systems has become essential to improve efficiency and enhance the attractiveness of rail transport. However, implementing AI in railway operations comes with several challenges. Ensuring safety is the top priority while simultaneously increasing throughput.
February 18, 2025
© iStock/hsvrs
Systems based on artificial intelligence (AI) can sometimes make decisions that are hard to understand, making it difficult for the public to trust them, especially in unexpected situations. The long timeframe railway systems will be in use adds to the challenges. Ongoing alignment on the scope of use among all stakeholders is necessary to ensure that safety requirements are consistently met. In this context, the compatibility and reusability of AI systems become crucial qualities.
AI solutions need to effectively address the needs of the context in which they are deployed. It is important to recognize when such AI systems are used beyond their intended purposes. Moreover, the quality and availability of data present significant hurdles. High-quality, relevant data is essential for training and validating AI models to ensure they cover both general and specific requirements and operational scenarios. It is important to establish agreement on the fundamental conditions for use within the domain, as well as the specific conditions of each project. Ensuring coverage of these conditions is necessary, and it is essential to determine when the system has departed from the agreed-upon operational conditions.
Operational Design Domain (ODD) is a key framework in this context, providing a clear structure for defining and managing the operational environments in which AI systems can function safely and effectively. ODD ensures that AI systems are designed with safety in mind from the start and remain reliable as operational demands change.
© Fraunhofer IKS
Figure 1: ODD enables traceability and consistency across different development disciplines
Safety is a fundamental aspect of any railway AI system and must be prioritized from the design phase. Compliance with industry standards and regulations is essential to ensure that AI systems meet necessary safety requirements. Traditional safety systems focus on identifying faults and minimizing risks, but AI introduces data-driven components that cannot always be removed from critical safety paths. This highlights the importance of carefully selecting data for training and testing AI models to accurately reflect the intended operational environment.
Clear, machine-readable definitions of operational environments are vital for effective AI system integration in railway settings. These definitions improve communication among stakeholders, align system capabilities with operational expectations, and guide comprehensive safety measures. Standards like EN 50126 and IEC 61508 require identifying environmental and operating conditions to ensure safety, performance, and reliability are built into the system design from the beginning.
Involving stakeholders throughout the development process promotes transparency and trust, aligning the system’s capabilities with user expectations and the environments in which it will operate. Continuous monitoring after deployment is crucial to detect and fix safety issues in real-time, ensuring the system remains safe during operation. Implementing fail-safe protocols allows AI systems to revert to a safe state if something goes wrong, maintaining safety in unexpected situations.
Defining operational environments also makes testing easier by providing relevant and complete testing scenarios. Thorough testing, including both simulations and real-world scenarios, is essential to identify and reduce potential risks before full deployment. High data representativeness, achieved through relevant and accurate data, improves decision-making reliability. Real-time monitoring benefits from these definitions by allowing the system to detect deviations from expected conditions and integrate fail-safe protocols when necessary.
An iterative development approach allows for continuous improvement by incorporating feedback and new insights, enhancing system capabilities while maintaining safety within the defined operational environments. Maintaining multiple views of operational environments—such as what is possible, expected, and currently supported—ensures compatibility and helps identify gaps, enabling focused enhancements. This comprehensive approach is captured in the concept of the Operational Design Domain (ODD).
Managing ODDs along the AI lifecycle easily becomes complex and becomes a critical development task. Thus, tool support and automation are crucial for effectively using ODDs in the iterative development of AI systems. At Fraunhofer IKS, the ODD Workbench was developed to help define and manage ODD taxonomies and definitions in a structured way. The workbench uses a custom model to ensure machine-readability, promoting a layered description of different aspects. A taxonomy layer can reference and refine other layers, allowing for a modular description. For example, one layer might cover general information about people, while another focuses on infrastructure buildings. A use-case-specific layer combines concepts from these layers into a unified taxonomy and adds attributes specific to particular applications.
© Fraunhofer IKS
Figure 2: Screenshot of the Workbench used to edit an ODD Taxonomy.
For modularity, the ODD Workbench uses two types of layers for ODD definitions: restriction and instance layers. Restriction layers use simple true/false statements to define what is suitable within an ODD, referencing specific attributes or constraints in the taxonomy. Instance layers clearly define the boundaries of relevant taxonomy attributes, allowing for uncertainty through ranges or option sets. This two-layer approach ensures that taxonomy references can accurately describe both expected conditions and actual measured scenarios.
© Fraunhofer IKS
Figure 3: Screenshot of the Workbench used to edit an ODD Description.
The workbench is user-friendly, using a textual Domain-Specific Language (DSL) that makes it easy to define and modify ODD taxonomies and definitions, similar to writing code. It supports versioning and version control, enabling collaboration and iterative development. Features like outline, auto-completion, verification, and syntax highlighting improve the user experience, while the ability to view and edit models with any text editor adds flexibility.
The ODD Workbench also includes tools for using and evaluating ODDs in real-world situations. These tools check how well ODD definitions match each other and existing test data. Mapping layers define the relationships between different taxonomies that may use different names or representations for the same concepts. This helps automatically check coverage and compatibility, connecting various taxonomies such as labels used in machine learning algorithms and customer expectations. For example, machine learning labels can be translated based on ODD concepts, generating Python scripts that automate the translation and make it easier for developers to integrate AI outputs with ODD specifications. Comparison tools ensure clarity and consistency across ODD taxonomies and definitions, while diagram generation provides visual representations of coverage, overlap, and gaps, simplifying the assessment of ODD adequacy and data set usage.
© Fraunhofer IKS
Figure 4: Screenshot of the Workbench mapping an ODD Taxonomy to ML Labels and visualization of the overlap.
Reporting tools in the workbench create detailed textual reports or CSV tables, helping with documentation and communication by making ODD taxonomies and definitions easily accessible, for example as an interactive webpage. Additionally, the workbench allows exporting ODDs in custom formats, enhancing usability and integration. This feature supports tailoring outputs to meet specific system needs and can be integrated into CI/CD pipelines, making ODD concepts available in existing lifecycle management tools and aligning closely with current practices in test planning and test-case generation.
© Fraunhofer IKS
Figure 5: Screenshot of an interactive webpage to browse ODD Taxonomies and Definitions.
© Fraunhofer IKS
Figure 6: Screenshot of a generated report for an ODD Taxonomy.
Integrating AI systems into railway operations offers significant benefits for improving efficiency and addressing personnel shortages. However, the successful deployment of these systems depends on clearly defining and managing their Operational Design Domains (ODD). Tools like the ODD Workbench from Fraunhofer IKS are essential for structuring and maintaining these domains, ensuring that AI systems operate safely, reliably, and effectively within their intended environments. As the railway industry continues to adopt AI, focusing on ODD will be crucial for building trust, ensuring safety, and driving continuous improvement in railway operations.
The project is funded by the European Union and the German Federal Ministry for Economic Affairs and Climate Action (BMWK) as part of the “New Vehicle and System Technologies” program.
