While LFV engineers worked on funding and data collection, IBM Garage experts from the Copenhagen, Denmark location pulled in the IBM Research® team and developers for the IBM® Streams analytics platform. The expanded LFV and IBM team collaborated during an IBM Enterprise Design Thinking Workshop™. Because the IBM Garage Methodology focuses on user-centered design, LFV included two air traffic controllers who provided insight about factors such as aircraft limitations, fuel load and pilot cooperation.

With a solid vision and technical plan, the team was ready to begin developing Advanced Autoplanner (AAP), an AI-driven autonomous air traffic control solution financed by the Swedish Transport Administration. But COVID-19 hit and the Nordic countries went into lockdown. Disruption is in the DNA of the agile IBM Garage Methodology, so the project remained on track. Team members across Europe and the US maintained a regular cadence of agile stand-ups, playbacks and technical status calls, leading to successful development of the first AAP minimum viable product (MVP).

Aviation standards require five nautical miles around each in-flight aircraft at all times. When building AAP, the LFV and IBM Garage team included a buffer, requiring six nautical miles. If planes are closer than this, it’s called a “loss of separation,” which—over time—can result in a collision. AAP operates in two phases as it oversees a specified airspace sector. First, a lattice-based 3D space exploration technique continuously tracks and forecasts aircraft locations in real time. If the app determines that a plane will experience a loss of separation, it is able to run nearly 800 possible scenarios—in one second—of slightly altering a plane’s direction, speed or altitude. AAP looks at how a scenario’s trajectory would affect the entire sector’s airspace, like a 3D chessboard, and then identifies safe actions that avoid future conflicts.

Second, the solution uses a rule-based approach to rank the actions identified in phase one and sends the best option to the pilot. The pilot can execute the instruction or communicate that it is not possible. For example, if the instruction is to increase the aircraft’s altitude by 1,000 feet and the pilot determines this is not feasible, the app would then provide an alternate instruction, such as adjust course five degrees east. AAP also tracks when the aircraft is safe to resume its original flight plan.

Surprisingly, the pandemic positively impacted a vital aspect of the project. Because LFV controls both civil and military air traffic in Sweden, its data is highly secure and therefore inaccessible offsite. During lockdown, the extended team was no longer permitted onsite to access data needed to build the AAP solution’s AI model. Forced to rethink AI development, the group decided to build an AI model based on integration with an NLR (Dutch National Aviation and Aerospace Laboratory) Air Traffic Control Research Simulator (NARSIM).

Data scientists built a deterministic AI model with a custom algorithm on the IBM Streams analytics platform and then opened a connection between the NARSIM simulator and IBM Streams running on IBM Cloud®. They simulated traffic across the determined airspace, and based on the output, continued to refine the algorithm and iterate the AI model. An IBM Db2® database stores data for the IBM Streams application and an IBM Cloudant® database stores the AAP solution’s instructions for the air traffic controller and pilot.

Headquartered in Norrköping, Sweden, LFV (link resides outside of ibm.com) provides air traffic control and associated services for civil and military aviation in Sweden. During normal situations (pre-pandemic), LFV air traffic controllers manage approximately 2,000 aircraft daily within the Swedish airspace. LFV employs 1,100 people and has an annual turnover of SEK 3.1 billion.