Data & Drinks van Kickstart AI
Chris Twigt, Competence Developer AI bij Pancompany, was onlangs aanwezig op het Data & Drinks event van Kickstart AI en Xomnia. Het thema AI & Aviation stond centraal en er waren sprekers van Schiphol & KLM. Chris schreef daar dit verslag over.
Last week I attended a “Data&Drinks” meetup organized by Kickstart AI and Xomnia about AI & Aviation (Schiphol and KLM). It was packed and I was fortunate to have a seat.
There where interesting talks about applying Data to reduce traveler queue length at Schiphol and applying Machine Learning to predict plane component failure. Here are the main take-aways – in bold:
Applying Data to reduce traveler queue length – Roman Kazus (Royal Schiphol Group)
When you intend to take a flight from Schiphol you can now book a “Security time slot” in advance – online. This way you are guaranteed to get through security in that time slot and you will bypass any of those dreaded security queues. The smart Data solution keeps track of queue-length peak times during the day to determine when passenger rate exceeds security personnel capacity. The Security time slots are subsequently offered on both sides of the peak times. This reduces the queue length by “flattening” the peak. So by booking a Security timeslot online for your flight from Schiphol you do yourself – and other passengers – a big favor!
Applying AI at Schiphol – Philip Kerbusch (Royal Schiphol Group)
Philip discussed applying machine learning in Schiphol in a general sense and addressing scheduling problems, in particular scheduling/assigning planes to gates.
They have constructed a solution using Reinforcement Learning which can instantly give good advice in case planes need to be rescheduled to different gates as a result of delays. However, the Schiphol personnel is still in charge. Autonomy – the highest achievable status for AI’s – will take more trust from us humans.
Predicting plane component failure using ML – Martijn Oerlemans & Dennis van den Berg (KLM)
KLM has developed multiple assistive approaches to determine component failure and preventive maintenance. Planes have a lot of sensors on board to track the health of the components. KLM receives an impressive 250GB of sensor data from their fleet every week!
They have developed a solution to predict component failure using ML that uses autoencoders for anomaly detection in sensor data. While an Autoencoder would probably do the job they took one additional step by mixing Transformers into the encoder and decoder of the AutoEncoder!
This is something completely new and original. Does it work? Yes! Remember the chain-rule for differentiation? All model components are differentiable so the model is too. Long live gradient descent!
