LESSONS LEARNED FROM HUMAN OPERATOR INTERVENTION FOR AI NAVIGATION AND FLIGHT MANAGEMENT SYSTEMS

The advent of Machine Learning (ML) has made possible the deployment of AI systems to perform complex decision-making formerly performed by "expert" human operators. Although modern airliner avionics were not developed using ML methods, they exhibit the characteristics of complex AI functions, especially regarding the complex decision making required for Flight Planning, Navigation, Guidance and Control functions. To address operational situations that the AI functions are not designed for, human operators are required to monitor and intervene should the AI functions command the vehicle into a hazardous operating region. Accident investigations of modern airliner accidents have cited the failure of the flight-crew to complete the required intervention Standard Operating Procedure in a timely manner. This will be an important issue that must be addressed in the design of the proposed new eVTOL/eSTOl aircraft. This paper describes a detailed analysis of the SOPs required for use in the operator intervention of the expert system AI functions involved in Part 121 airliner and Part 135 accidents. The analysis identified two main categories of SOP design issues regarding the time it takes to complete the intervention task: (1) designed SOP takes a longer time than the actual allowable time, i.e., it is humanly impossible to complete the intervention task in time and ( 2) the designed SOP ToP takes reasonable time, but in rare cases the actual time to complete the intervention task is greater than the actual allowable time due to execution problems. The implications of these results are discussed.