Towards Comprehensive Safety Assurance of a DAL A AI/ML-based Runway Alignment System using Overarching Properties

Details

• Creators:
  Paul, Saswata ; Prince, Daniel ; Iyer, Naresh ; Tang, Liang ; Durling, Michael ; Meiners, Mike ; Meng, Baoluo ; Chandra Varanasi, Sarat ; Visnevski, Nikita
• Corporate Creators:
  GE Aerospace Research ; GE Aerospace ; GE Vernova Advanced Research
• Corporate Contributors:
  United States. Department of Transportation. Federal Aviation Administration. Senior Technical Experts Office (AIR-20)
• Subject/TRT Terms:
  AI/ML Airport Runways Aviation Safety Landing Fields Overarching Properties Runway Alignment System Safety Analysis
• Publication/ Report Number:
  DOT/FAA/TC- 26/20
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States
• Edition:
  Final Report
• Corporate Publisher:
  United States. Department of Transportation. Federal Aviation Administration. William J. Hughes Technical Center
• Abstract:
  This report details an Overarching Properties (OPs)-based approach for assuring the safety of Artificial Intelligence/Machine Learning (AI/ML)-based digital aerospace systems. To rigorously evaluate and enhance this approach, an AI-Assisted Autonomous Runway Alignment (AARA) system is introduced as a motivating use case, allowing for the identification and mitigation of potential safety risks through premise-based arguments. The study demonstrates how multi-level safety assessments can be conducted for AI/ML systems to pinpoint failure conditions inherent to AI/ML's nature. It also provides examples of how requirements can be formulated to address risks posed by black-box AI/ML components with unpredictable or uncontrollable behaviors. Comprehensive discussions cover various aspects of the OPs-based approach, including foreseeable operating conditions, development and training activities, evidence generation for premises, hybrid certification, design assurance, necessary assumptions, and a plan for OPs compliance. While currently focused on Artificial Neural Networks (ANNs) developed using supervised learning, the arguments may be adaptable to other AI techniques. The findings establish a strong foundation for applying OPs to AI/ML assurance, acknowledging the need for further work to ensure robustness and practicality across diverse criticality and autonomy spectrums.
• Format:
  PDF
• Collection(s):
  FAA William J. Hughes Technical Center for Advanced Aerospace
• Main Document Checksum:
  urn:sha-512:a3693102f916afc4bc22f1bac2c23bae4243c9a976546ca71765f76ca51da2613c0eeccac618983f6849a9449896b82c576e8b68404bf16a7dad18c48c1680bc
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/92476/dot_92476_DS1.pdf
• File Type:
  [PDF - 3.53 MB ]