Mastering Training Data Generation for AI - Integrating High- Fidelity Component Models with Standard Flight Simulator Software

The German state-funded aviation research project “Real- time Analytics and Prognostic Health Management” (RTAPHM) envisioned fully automated urban air services executed by autonomous drones and infrastructure controlled by a digital system. Research was focused on utilizing onboard real-time diagnostics to enable AI-driven UAV capability predictions. These predictions increased the reliability of upfront service commitments. The use case selected to demonstrate these elements was organ transport. The project delivered an end-to-end demonstrator incorporating a virtual fleet of drones with onboard diagnostics to provide data for the platform decision logic. The project followed a „digital-twin-first” approach to overcome a common bootstrapping problem faced by data- driven applications. That is, the lack of in-service data for exploration, prototyping and training of diagnostic and prognostic approaches during the concept and early development phases. Due to the upfront development of physical high-fidelity simulation models for the monitored components, a digital twin – of the portion of the twin that resembles the physical behavior – was used to generate data and facilitate preliminary exploration, prototyping and training. Digital twins were further employed to allow evaluation of what-if scenarios and identify the optimal future operation parameters of a drone. Development of the RTAPHM digital twin involved a multi- disciplinary team of members distributed across different organizations and locations. Successful realization of the digital twin depended on early integration testing, performed in high frequencies, which generated continuous feedback regarding technical and conceptual issues. Within the research project we developed MOLE, an engineering tool for automating the integration of distinct simulation components, into a single system simulation driven by commercially available flight simulator software. Here, we showcase the internal mechanisms of the tool and demonstrate its abilities to generate a Docker-based executable for efficient data generation in the cloud. We also show our approach to online visualization, fault insertion, batch integration testing and debugging the digital twin executable. We also report on the utilization of MOLE in assembling the final RTAPHM demonstrator