Dependable Deep Learning for Safety-Critical Airborne Embedded Systems
|Coordinator||SAAB Aktiebolag - Saab AB, Avionics Systems|
|Funding from Vinnova||SEK 3 067 606|
|Project duration||September 2019 - December 2022|
|Venture||National Aeronautical Research Program 7|
|Call||Research project in aviation technology - spring 2019|
Purpose and goal
This application addresses design methods for the use of deep neural networks in airborne safety-critical systems. When designing such systems traditional design assurance methods detailed in existing certification agency documents cannot be applied. The research focus of the project is on design error mitigation both in hardware and software as well as for other types of errors, e.g. contradictive, which could lead the system to express erroneous and unintentional behaviour.
Expected results and effects
The project investigates and evaluates methods and processes for the introduction and use of safety-critical deep learning systems. The expected outcome includes design methods and fault-tolerant architectures for airborne safety-critical systems implementing deep neural networks. The project is expected to produce a demonstrator for such a deep learning fault-tolerant architecture.
Planned approach and implementation
The project is split into four work packages. The first runs throughout the project and has a coordinating role. The second package will study the use of ML in airborne safety-critical applications and ML-based fault-tolerant architectures. The third package aims to evaluate design assurance methods and define a fault-tolerant architecture for a specific use-case. The final package will disseminate the results on international conferences. Important conclusions for the industry partner will be included in a lessons learned report which will be produced.