Hacker News Reader

Comment by oron | Hacker News Reader

At SkyPath we have a simple solution that doesn't involve installing any HW on the aircraft and uses the airline issued iPad the pilot already uses for other important tasks to help fight Clear Air Turbulence. We use the iPad's accelerometers to measure and report turbulence in real time and collect this data via the internet where it's processed in AWS by our machine learning model to produce customized alerts for each flight that uses our service on incoming CATs in their route.

View on HN

amazing! it seems like much of the innovation on the flight deck is happening in iPad apps.

i'm curious, in the beginning, before you had so many users of the app, how did you convince pilots/airlines to install SkyPath before it had enough user-generated turbulence data for its ML model? it almost feels like a chicken-and-the-egg problem: you need enough reports before it's useful around the world at all imaginable air routes, or maybe there's enough air/wind data. interesting stuff!

We had to give incentives for the first airline partners and give the product for free for a long trial period at the early years to be able to have initial installs. It was a long ride including Covid which came in the middle and didn't help. In the last years since we have several big airline partners this is less of a problem.

The effect on the route is immediate. There is no installation or integration so the moment you deploy (with a reasonable fleet size) you have data. Those that where quick to understand the concept embraced technology very fast even without initial data set in the route

Didn’t ForeFlight just come out with something similar recently? https://blog.foreflight.com/2024/05/07/smooth-skies/ They don’t mention ML specifically, but I imagine their install base is a tad larger and can therefore gather more real data. Is your solution different?

We have a partnership with Jeppesen, a Boeing company, for their Flight Deck Pro product. With this collaboration, users can access our data layer and predictions, and also report real-time turbulence within their app. While I'm not extensively familiar with the inner workings of ForeFlight, based on what I've heard, it relies on an external hardware solution, and its coverage and quality are not comparable

Note that Foreflight is a Boeing product. Therefore they could potentially have access to actual real-time aircraft data that this product would not.

Our solution is compatible with Airbus, Boeing, or any other aircraft. It's important to note that connecting to the aircraft systems can be more expensive and add maintenance and complexity costs for the airline.

What goes into developing, testing, calibrating something like this?

Apart from usual SW tests, Lots of testing with airlines and experienced pilots. Processing feedback and improving the model step by step over the course of several years.

I'm very curious how you "bootstrap" a model like that. Do you start with a physics simulation? Make an educated guess and then get pilots to label when it's wrong in order to incrementally improve? Recruit pilots to manually track CAT events?

I've had to build a few of these kinds of models over the years, and it's consistently the hardest task I've faced in my DS career.

The second is closer to reality , initially let pilots (lots of them) label flight events and after landing take all data and build the algorithm around their initial labels. The when there are more pilots let them label agree / disagree and relabel etc.

Another aspect is you have usually two iPads in the cockpit for captain and first officer so you can correlate and match what both accelerometers read.

Thanks! It's nice validation to know that my approach to this kind of problem similar to that of a successful AI startup.

Ipad accelerometers and GPS

Does your app have to run in the foreground on the iPad during the flight? Or can the crew user other apps in flight?

Can run in the background, pilot can use any other app meanwhile. Or just leave the iPad with screen off. Our app keeps recording and alerting as long as you are in flight.

I'm assuming that the most valuable data for predicting turbulence during a flight is timely data from the accelerometers of the iPads on nearby airliners, which gets me curious as to whether these iPads usually have connectivity (to your AWS servers) during a flight, and how that connectivity is provided. (I'm guessing satellite.)

Yes, satellite internet is now standard on many Western airlines. In addition to this, we utilize weather data and real-time turbulence reports to predict Clear Air Turbulence (CAT) events across the entire sky, even where iPads are currently not on flights.

how exactly are you using an accelerometer to determine turbulence

By analyzing the acceleration patterns that the iPad the pilot has in the cockpit which is securely attached to the aircraft is measuring. Cleaning out noise such as engine and other non CAT noise and training our model to identify these and also predict future events in the next 3 hours by looking at the current data. It takes millions of events and lots of pilots which help train the initial model and fine tune it.

How else would you detect it? Accelerometers seem perfect for this use case.

Reply on news.ycombinator.com