Learning a better phone vario from real flight data

I am an AI modeler at heart, and after starting paragliding I've thought that it would be cool to make a better vario and wind estimator based on real flight data. However, I've always had two issues: I haven't had the correct data, and I've had no place to deploy a smarter thermal algorithm.

But now I do!

It's early stages, but now that people are starting to fly with pgpilot I am able to start experimenting with different sensors and data. For example here, where I collect barometric data from three different sensors during flight (Enable "benchmark logging" in hardware section to contribute!).

The model is a Kalman-filter, previously hand tuned, now learnt from real data to de-smooth the noisy barometer signals.

And it's already paying off: training on barometric and fusing in accelerometer allows me to massively reduce the previous 1sec lag for the internal phone vario. So that is the new algorithm in pgpilot in the release that arrived this morning. Not as good as an external vario, but good for a hike&fly or if you forget.

The image shows data from my phone, a VectorVario and an XCTracer barometer over time. It's just a tiny dataset, but I hope we can collect more data soon so we can both benchmark various varios AND improve the overall experience. For example, would we be able to make an even better vario & thermal sniffer by fusing in both phone and external vario+accelerometer into one?

Now it's time to collect some data, hoping for good weather soon!