About the project

WP1: Requirements for the Safety Case Verification 

The purpose of the work package is to research and survey the existing standards in terms of applicability to AI-based and ML-based safety functions, derivable requirements for and properties of such safety functions, and the methods, tools and metrics to be used. 
 
As a result, this work package will provide requirements and guidelines for demonstrating the product safety of a GoA4 rail vehicle with AI-based safety-related functionalities for certification purposes. 
 
To this end, the work package is subdivided into the following topic areas: 
 
Applicability analysis of existing standards 
All existing standards and norms from the areas of rail, industry, automotive, security, functional safety and AI in general will be recorded and examined in order to ascertain the extent to which they can also be applied to new, safety-critical AI systems. 
 
The focus here is on researching relevant activities and existing norms, standards, harmonized norms under the NLF and CSA schemes, as well as best practices and experiences with relevant AI applications. The final step will be to derive the applicability and the limits of existing norms and standards, for example from the abovementioned areas. 
 
Derivation of properties and acceptance criteria to be demonstrated for the safety functions 
The basic properties of and requirements to be met by AI-based systems for automatic train control must be suitable for ensuring a level of safety which is equivalent to the classic approach for assessing systematic faults (e.g. Safety Integrity Levels according to IEC 61508/EN 50129). 
 
The derived properties and acceptance criteria to be demonstrated will form the basis for the subsequent verification and validation of the safety of the AI-based application for automatic train control. 
 
Requirements to be met by methods, tools, development processes and operation 
On the one hand, requirements will be drawn up for methods and tools for the verification, validation, testing and assessment of AI systems. In the specific use case, the focus here is on perceptive AI applications based on neural networks. On the other hand, requirements will be determined for processes for developing and operating AI systems. 
 
Quantifiable metrics for assessing the trustworthiness, declarability and safety performance of AI 
Based on the previously defined properties and acceptance criteria to be demonstrated and the results of the applicability analyses of existing norms and standards, state-of-the-art metrics will be researched and analyzed in order to provide a quantitative assessment of system-side safety objectives and AI-specific properties.  
 
Here, the focus is on defining requirements and metrics in order to be able to assess objectively whether the acceptance criteria for the certification of an AI function are met in the proposed use case.

WP2 Methods and Tools for Establishing and Demonstrating the Trustworthiness of AI Functions 

The objective of this work package encompasses two central topic areas. The first relates to the systematic creation of safety requirements for the AI-based systems used within the project. The second area focuses on the creation of methods and proofs that can be used to validate the AI requirements defined in the first area. 

Considered in more detail, the two areas of the work package comprise the following points: 

WP3: Safety Architecture for AI-based Functions in GoA4 Operation 

This work package covers the development of an integrated basis for requirements and architecture as well as overall system implementation (system under test) and the safety case for the AI-based object detection of a driverless regional train. 

The work package is subdivided into 4 topic areas: 

Requirements 

Based on results from WP1 and findings from research projects at the national level (e.g. ATO-Risk, ATO-Sense) and at the European level (e.g. X2Rail, TAURO), the requirements to be met by a driverless regional train with an AI-based object detection system will be collected, prioritized, and integrated into an overall structure. This will be supplemented by the definition of the operational design domain, in which the environmental conditions (operational, climatic and weather-related constraints), including the persons, obstacles and systems interacting with the system, are recorded. 

Architecture 

Based on the defined requirements, an architecture for a "driverless regional train" will be drawn up for the object detection system, the sensor fusion and the sensor system, in which the interfaces between the functions are defined at all 4 architecture levels and the RAMS requirements (reliability, availability, maintainability and safety) as well as NFRs (non-functional requirements) are taken into consideration. 

Implementation 

In the third subarea of work package 3, the overall system will be implemented and integrated as a software solution. The ML functions/models provided for in the architecture will be implemented here. These form the basis for the methods and metrics developed in WP2 for investigating the trustworthiness and validation. In addition, a hardware demonstrator will be developed and tested for the sensor system. 

Safety case procedure 

The final subarea covers the development of a concept for a safety case for a driverless regional train with an AI-based object detection system, combining the results of the upstream activities in the area of requirements, architecture and metrics. The structure of the safety case serves as the basis for evaluation in the virtual test environment (WP4) and subsequent assessment.

WP4: Virtual Test Environment, Safety Analysis 

WP4 covers the development of a virtual test environment for evaluating the project objectives in safe.trAIn. In this virtual test environment, the validation of an AI function in the rail transport sector can be carried out using the example of obstacle detection. 

The work package is subdivided into the following topic areas: 
 
SafeMLOps 
Based on the results of work packages 1-3, an end-to-end development infrastructure for AI methods for rail vehicle applications will be designed and implemented as a prototype into which the investigated methods are integrated. To this end, the test methods required to generate proofs (defined in WP2) for the safety case procedure will be integrated into what is known as a SafeMLOps process and automated. 
 
Virtual test environment 
In order to implement this SafeMLOps approach as a prototype, a virtual test environment for driverless vehicles will be developed in WP4 for the purpose of validating the test criteria for AI functions. This virtual test environment will serve as a validation platform, integrating the different test methods developed in WP2 and implemented in WP3 in order to verify the trustworthiness of an AI function and generate proofs for the safety case procedure. Here, the AI function will be validated by means of software-in-the-loop methods with the aim of carrying out comprehensive empirical tests (simulations, etc.) with the AI functions and obtaining metrics for the safety case procedure. These simulations will be carried out with both real and synthetic data. 
 
Evaluation 
The results from the virtual test environment will then be compared with the defined test criteria of the certification process and embedded in a safety case for driverless rail vehicles. This will enable an initial set of rules to be tested under real conditions, from assessment through to conformity declaration for the present use case, and will allow conclusions to be drawn as to the applicability of test criteria and test methods in the certification process. The results obtained here will be forwarded continuously to WP1, WP2 and WP3.

WP5: Standardization and Dissemination

Work package 5 deals with the standardization and distribution of relevant results from the safe.trAIn project. This will facilitate a broad application of the results even after the end of the project. 
 
The work package is divided into the following focal topics: 
 
Standardization activities 
To ensure the successful transfer of results into standardization activities, various tasks are addressed in the corresponding work package sections. The essential tasks are to conduct an analysis of the standards identified in work package 1 as relevant to the project and to identify where there is a need for standardization, e.g. standardized interfaces, processes, test procedures, etc. These will be transferred to standardization activities in close collaboration with all project partners and relevant standardization committees. The aim is to transfer relevant project results into standardization with the involvement of other stakeholder groups. There is also a focus on ensuring the most strategically effective distribution of relevant project results at the national, European and international standardization levels. 
 
Formation of user groups 
Another task is to set up a networking and discussion platform. This will take the form of user groups, with the aim of validating project results with other expert stakeholders, considering use cases and where appropriate identifying joint standardization potential.  
 
Consideration and analysis of the AI ecosystem in terms of normative legislation  
In view of the changing legislative framework for normative requirements in the world of artificial intelligence, the transfer of project results (e.g. validation argumentation) to established dimensions for trustworthy AI (e.g. data protection, fairness, transparency, autonomy and control, technical documentation, robustness, reliability, safety, etc.) will be investigated. In addition, a methodology will be developed for the fulfillment of all regulatory requirements, as stipulated for example in the current draft of the European AI regulatory framework. This will be supplemented by an analysis of the project-specific effects of further regulatory developments (e.g. Cyber Resilience Act). 
 
Distribution of results 
The project results are also to be published outside the safe.trAIn project. To this end, it will be necessary to define measures in order to ensure that the project activities are communicated effectively, and to support this process.