Ensuring Safety of Machine Learning Components Using Operational Design Domain

Torens, Christoph; Juenger, Franz; Schirmer, Sebastian; Schopferer, Simon; Zhukov, Dmytro; Dauer, Johann C.

doi:10.2514/6.2023-1124

Cited by 3 publications

(1 citation statement)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Basically, this requires monitoring the input and output of the ML component for unsafe actions in order to mitigate risk, similar to recovery control function in ASTM F3269-21. Further details on the analysis of monitoring aspects in the standard appear in [11]. ASTM F3269-21 not only describes a reference architecture for assuring the safety of black-box systems, or systems External Data Safety Monitor Complex Function Recovery Control Function(s) Switch Fig.…”

mentioning

confidence: 99%

A Hierarchy of Monitoring Properties for Autonomous Systems

Schirmer

Torens

Dauer

et al. 2023

AIAA SCITECH 2023 Forum

View full text Add to dashboard Cite

Monitoring capabilities play a central role in mitigating safety risks of current, and especially future autonomous aircraft systems. These future systems are likely to include complex components such as neural networks for environment perception, which pose a challenge for current verification approaches; they are considered as black-box components. To assure that these black-boxes comply with their specification, they must be monitored to detect violations during execution with respect to their input and output behaviors. Such behavioral properties often include more complex aspects such as temporal or spatial notions. The outputs can also be compared to data from other assured sensors or components of the aircraft, making monitoring an integral part of the system, which ideally has access to all available resources to assess the overall health of the operation. Current approaches using handwritten code for monitoring functions run the risk of not being able to keep up with these challenges. Therefore, in this paper, we present a hierarchy of monitoring properties that provides a perspective for overall health. We also present a categorization of monitoring properties and show how different monitoring specification languages can be used for formalization. These monitoring languages represent a higher abstraction of general-purpose code and are therefore more compact and easier for a user to write and read, and we can validate their implementations independently from the systems they reason about. They improve the maintainability of monitoring properties that is required to handle the increased complexity of future autonomous aircraft systems.applied monitoring based on specification languages. Next, we present a hierarchy of monitoring properties that ranges from low-level sensor properties to high-level operation properties, giving a holistic perspective on possibilities for system monitoring. Further, we categorize different types of monitoring properties and showcase different monitoring specification languages. Finally, we discuss the advantages and disadvantage of using such formal languages. Related Work: Aerospace RegulationsAviation standards and regulations offer a perspective on target systems that informs their monitoring. For instance, Aerospace Recommended Practice (ARP) 4754A [3] from SAE International is the guidelines document for development of civil aircraft and systems. Additionally, SAE International's ARP4761 [4] provides guidelines for conducting the safety assessment process on civil airborne systems and equipment. The terms used in both documents for developing an aircraft as well as performing a safety assessment on the different process levels are item, system, and aircraft. An item, is one or more hardware and/or software elements treated as a unit. An item is the lowest level of abstraction during development. A system is a combination of inter-related items arranged to perform a specific function. A system therefore represents a higher abstraction level than an item. The aircraft is...

show abstract

mentioning

confidence: 99%

A Hierarchy of Monitoring Properties for Autonomous Systems

Schirmer

Torens

Dauer

et al. 2023

AIAA SCITECH 2023 Forum

View full text Add to dashboard Cite

show abstract

HorizonUAM: operational challenges and necessary frameworks to ensure safe and efficient vertidrome operations

Schweiger,

König,

Metz

et al. 2024

CEAS Aeronaut J

View full text Add to dashboard Cite

Urban Air Mobility (UAM) has emerged as a potential game changer for urban transportation, promising faster, more efficient and affordable services. However, beyond the visionary concepts, it is crucial to explore and discuss the opportunities and challenges of UAM and vertidrome operations also from a research perspective. The DLR research project HorizonUAM aimed at a holistic research approach in which vertidromes and vertidrome networks play a significant role. This vertidrome centered project report covers various aspects and methodological approaches addressing design and operation, UAM airspace management, network optimization and the integration of air taxi operations into airport environment. Moreover, the conceptual and temporary development of a modular 1:4 scale model city lays the foundation for future UAM flight trials. Based on three years of dedicated research within HorizonUAM, we focus here on operational challenges, proposed solutions and required frameworks to ensure safe and efficient vertidrome operations

show abstract